🤖 Robotics & Embodied AI¶
🧠 NeurIPS2026 · 132 paper notes
📌 Same area in other venues: 🎞️ ECCV2026 (135) · 📷 CVPR2026 (163) · 🔬 ICLR2026 (166) · 💬 ACL2026 (11) · 🧪 ICML2026 (54) · 🤖 AAAI2026 (34)
🔥 Top topics: Robotics ×35 · Multimodal/VLM ×23 · Agents ×15 · Navigation ×6 · Alignment/RLHF ×5
- A2Eval: Agentic and Automated Evaluation for Embodied Brain
-
A2Eval is the first agentic evaluation framework for embodied VLMs that uses a Data Agent to induce an 8-dimensional capability taxonomy and curate a balanced compact suite via diversity-aware clustering, while an Eval Agent synthesizes and validates executable evaluation code, compressing benchmark volume by 85% and speeding up evaluation by 4.6× while correcting leaderboard ranking distortion.
- A latent control model for realistic rodent motion
-
To resolve the trade-off between realism and diversity in embodied animal reinforcement learning, this paper proposes SCAMPER, a dual-prior framework that combines a motion prior for goal-directed control with a behavior prior for exploratory sampling, enabling realistic movement during both goal tracking and sparse-reward exploration.
- Action Chunking Proximal Policy Optimization with Feedback Correction
-
ACPPO-Corr extends PPO with a low-frequency action-chunk planner and a stepwise feedback corrector while retaining a state-only value network, improving final normalized interquartile mean over PPO by 30.4% across 25 simulated robotics tasks—not by 30.4 percentage points of success rate.
- Action Images: End-to-End Policy Learning via Multiview Video Generation
-
Translating 7-DoF robot actions into multi-view pixel-grounded Gaussian action images enables a unified Wan 2.2 video generative backbone to serve directly as a zero-shot manipulation policy without separate action heads.
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning
-
ActionUNet constructs a lightweight temporal U-Net in the action feature space to inject hierarchical temporal priors and applies a conditional SIREN continuous decoder with explicit Matérn second-order smoothness constraints, substantially improving VLA robustness against environmental perturbations with minimal overhead.
- AdaOcc: Adaptive 3D Occupancy Prediction for Embodied Tasks
-
AdaOcc proposes an adaptive 3D semantic occupancy prediction framework for embodied tasks that unifies heterogeneous multi-sensor inputs via geometry-guided dual-branch encoding, enables budget-scalable inference through progressive query learning, and enforces high boundary fidelity with a novel containment loss.
- AnchorWorld: Embodied Egocentric World Simulation with View-Based Evolution Customization
-
AnchorWorld integrates hybrid-view projection-based motion control with pose-associated anchor views on top of the Wan2.2 video generation backbone, enabling full-body motion-driven egocentric world simulation with spatially grounded, text-guided local scene state evolution.
- APIVOT: Adaptive Planning with Interleaved Vision-Language Thoughts
-
APIVOT is a VLM-based planner that interleaves language with latent visual thoughts within its autoregressive reasoning trace to internally verify geometric feasibility during planning, significantly outperforming general-purpose VLMs and traditional planning frameworks on complex long-horizon kitchen manipulation tasks.
- ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection
-
Addressing the observation-action modality gap and joint optimization conflicts in predictive vision-language-action models, ATI-VLA introduces a two-stage "Align-Then-Inject" framework that anchors predictive representations into an actionable discrete latent space and adaptively injects them into action decoding under a pure action-centric objective.
- BEACON: Cross-Domain Co-Training of Generative Robot Policies via Best-Effort Adaptation
-
BEACON casts cross-domain robot policy co-training into a discrepancy-aware importance-reweighting problem derived directly from target-domain generalization bounds, significantly improving sim-to-real transfer and out-of-distribution robustness without any explicit feature alignment objective.
- Behavioral Foundation Models for Quality Diversity
-
BFM-QD freezes an offline-pretrained behavioral foundation model, builds a repertoire of high-quality, diverse behaviors by searching latent codes rather than policy weights, and uses closed-form backward inference for small directed mutations, substantially improving sparse navigation and contact-rich manipulation while still requiring environment rollouts and pretraining investment.
- Beyond Appearance Shifts: Task-Semantic Action Calibration for VLA Models
-
Addressing the dual failure modes where vision-language-action (VLA) policies drift under task-preserving appearance shifts yet obstinately follow stale trajectories under semantic-breaking changes, BAS-VLA introduces a lightweight calibration layer atop a frozen base VLA that combines a breaking-centered residual core with an evidence-gated preserving auxiliary, boosting changed-task completion from 2.3% to 68.5% while suppressing stale behavior.
- Change-Robust Online Topological Memory for Long-Term Relocalization and Semantic Navigation
-
Addressing the vulnerability of traditional spatial memories to appearance changes and false loop closures, CROSS introduces a pre-commitment localization layer between visual place recognition and map update that validates competing continuous \(\mathrm{SE}(3)\) pose hypotheses against odometry within a bounded Gaussian-mixture filter, significantly improving long-term spatial reuse and object-goal navigation.
- CLoSeR: Closing the Loop for Long-Context Streaming Reconstruction
-
Addressing the severe drift accumulation in feedforward streaming 3D reconstruction models over long trajectories, CLoSeR leverages the global scale consistency and non-contiguous inference capabilities of streaming backbones to construct loop-conditioned windows and optimize poses directly on the \(\mathrm{SE}(3)\) manifold, achieving drift-free kilometer-scale monocular reconstruction.
- Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning
-
To overcome global incoherence when composing short-horizon diffusion priors for long-horizon tasks, CoFi introduces an inference-time sampler that first establishes a shared coarse global scaffold and then refines local details via structure-preserving re-diffusion, cutting denoiser evaluations by 2–8×.
- CofactVLA: Deconfounding Vision-Language-Action Models via Counterfactual Intervention
-
Addressing the "vision-override" failure mode where dense visual observations bypass sparse language instructions in Vision-Language-Action (VLA) models, CofactVLA formulates a Dual-path Deconfounding Graph that isolates visual shortcuts via a language-masked counterfactual branch within a single forward pass, eliminating visual confounders at both feature and action levels to achieve a 52.3% gain in real-world OOD manipulation.
- CofactVLA: Deconfounding Vision-Language-Action Models via Counterfactual Interventions
-
To alleviate the "vision-override" phenomenon where dense visual streams dominate sparse language instructions in Vision-Language-Action models, CofactVLA formalizes action generation as a Dual-path Deconfounding Graph, eliminating collinear visual bias via Action-Level Orthogonal Projection Guidance and suppressing visual confounders via Feature-Level Counterfactual Covariance Reduction, delivering a +52.3% absolute success rate gain under real-world out-of-distribution scenarios.
- Continually Evolving Skill Knowledge in Vision Language Action Model
-
Stellar VLA enables parameter-growth-free continual imitation learning for VLA models by co-evolving Dirichlet-Process-based task/skill knowledge spaces and guiding MoE expert routing via knowledge posteriors, achieving state-of-the-art manipulation performance with only 1% replay.
- CrafterDojo: A Suite of Foundation Models for Building Open-Ended Embodied Agents in Crafter
-
CrafterDojo overcomes the lack of foundation models and demonstration data in the lightweight 2D Crafter benchmark by providing behavioral prior C-VPT, vision-language grounding C-CLIP, and instruction-following C-Steve-1, accompanied by automated data generation toolkits that enable fast prototyping and hierarchical planning without Minecraft's heavy engineering overhead.
- CSO: Refining Robotic Policies via Skill Distribution Alignment and Skill-Grained Optimization
-
Addressing quantization error accumulation in imitation-learned discrete skill policies as well as encoder instability and optimization granularity mismatches in online reinforcement learning, CSO proposes a two-stage post-training framework combining Rejection-Sampling Supervised Fine-tuning (R-SFT) and Skills Policy Optimization (SPO), boosting LIBERO overall success from 80.1% to 97.5% and achieving 61.0% success across physical robot tasks.
- DeformGen: Dynamics-Based Topology Augmentation for Deformable Manipulation Policy Learning
-
Addressing the failure of rigid SE(3) data augmentation on deformable manipulation, DeformGen introduces forward physics-based localized perturbation and deformation-field trajectory warping to synthesize physically plausible, topologically diverse demonstrations from a single demonstration, substantially boosting policy generalization.
- DeformMaster: An Interactive Physics-Neural World Model for Deformable Objects from Videos
-
DeformMaster transforms phone-captured or multi-view interaction videos into an online interactive physics-neural world model, combining differentiable Material Point Method (MPM) rollouts with neural velocity residuals, distributed compliant actuation, and a mixture of constitutive experts to drive high-fidelity 4D Gaussian appearance under novel interactions.
- Differentiable Learning of Lifted Action Schemas for Classical Planning
-
Introduces DIAS (Differentiable Induction of Action Schemas), an end-to-end differentiable architecture that recovers lifted STRIPS action schemas from state traces with unobserved action arguments via R-GCN object embeddings, log-space rectangular Sinkhorn slot allocation, and smooth precondition/effect projections optimized through gradient descent.
- Differentiable Nonlinear Model Predictive Control
-
This paper presents an efficient framework for computing forward and adjoint solution sensitivities of general nonlinear optimal control problems using the implicit function theorem and interior-point smoothing, featuring a two-solver architecture implemented in acados that outperforms state-of-the-art differentiable solvers by over 3x on CPU.
- DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising
-
Addressing the issue where intermediate projection overconstrains noisy latents and corrupts diffusion dynamics, DiRecT formulates safe planning as a terminally constrained stochastic optimal control problem and solves it via a receding-horizon data-domain surrogate without retraining.
- DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control
-
DiT4DiT introduces an end-to-end dual Diffusion Transformer architecture that jointly optimizes video dynamics and continuous action prediction via an asymmetric tri-timestep dual flow-matching objective, achieving state-of-the-art success rates on LIBERO (98.6%) and RoboCasa-GR1 (50.8%) while delivering a 10x sample efficiency boost.
- Done, But Not Sure: Disentangling World Completion from Self-Termination in Embodied Agents
-
Addressing standard embodied benchmarks that conflate execution failures with termination failures, this paper introduces Vigil, an evaluation framework under a strict no-feedback native-control protocol with mandatory semantic self-reports that disentangles world-state completion from benchmark success, exposing sharp behavioral divergence across embodied agents in terminal commitment.
- Doob-Based Generative Policy Optimization
-
Addressing the intractable action likelihood bottleneck in RL fine-tuning of diffusion and flow policies, this paper decouples policy improvement into lightweight ratio estimation and Doob h-transform-guided conditional drift matching, unifying PPO and GRPO while provably achieving minimum-energy stochastic optimal control.
- DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation
-
DynaFLIP pushes motion understanding upstream into robot visual perception by learning from image transition, language, and 3D flow triplets via simplex volume minimization with cosine regularization and contrastive learning, yielding a highly transferable single-image dynamics-aware visual encoder.
- DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation
-
To resolve the latency-induced perception–execution mismatch in dynamic manipulation where moving object states evolve during inference, DynamicVLA introduces a compact 0.4B architecture paired with continuous non-blocking inference and Latency-aware Action Streaming (LAAS), supported by the large-scale DOM benchmark spanning 200K synthetic and 2K teleoperation-free real-world episodes.
- EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields
-
EA-WM projects low-dimensional robot kinematic states into target-view Structured Kinematic-to-Visual Action Fields (KVAFs) and leverages Event-Difference Latent Supervision (EDLS) to dynamically gate bidirectional cross-attention, resolving domain misalignment and accurately preserving physical interaction dynamics in generated robot rollouts.
- Efficient Test-Time Adaptation For Robot Policies
-
Addressing the catastrophic failure of reinforcement learning policies deployed under unmodeled real-world distribution shifts, TEMPO introduces a lightweight on-device test-time adaptation framework that guides online actor updates via a frozen pretrained critic, regularizing critic predictive entropy and prioritizing low-value states in memory without requiring environmental rewards.
- EgoTac: In-the-wild Tactile Prediction from Egocentric Vision
-
EgoTac establishes the first 5.7M-sample visual-tactile corpus unified on a shared bimanual MANO topology and presents a transformer model combining temporal cross-attention and AdaLN modulation, enabling zero-shot, full-hand continuous force and contact prediction directly from egocentric RGB video.
- Embodied Neurocomputation: A Framework for Interfacing Biological Neural Cultures with Scaled Task-Driven Validation
-
This paper introduces the Embodied Neurocomputation systems framework, formalizing bidirectional bio-silicon interfacing as a multi-variable optimization problem and screening approximately 1,300 encoding configurations over 4,000 hours of closed-loop navigation across 26 biological neural cultures to significantly outperform silicon-based DQN baselines under equivalent interaction budgets.
- Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models
-
Embodied-R1.5 unifies spatial cognition, task planning and correction, and embodied pointing within a single 8B foundation model, leveraging balanced multi-task reinforced fine-tuning and a Planner-Grounder-Corrector (PGC) closed-loop framework to achieve autonomous long-horizon manipulation and error recovery without large-scale action pretraining.
- EmbodiedWorld-200K: Toward Embodied World Agents via Embodied-Planning Dataset and Interactive World Models
-
Addressing the generalization bottleneck of embodied agents in open worlds and the high cost of real interactive testbeds, this paper introduces EmbodiedWorld-200K—a 200K+ trajectory dataset with multi-level annotations from 150+ games—and proposes a two-stage training recipe combining offline supervised pretraining with video world model-based interactive RL post-training (DPO) to significantly boost open-world navigation-planning.
- Emergence of Physical Intelligence via Controllable Information Production
-
This paper introduces Controllable Information Production (CIP), an intrinsic motivation objective defined by the sum of positive open-loop Lyapunov exponents derived directly from the optimal control value function curvature, which drives agents to discover and stabilize complex physical behaviors such as self-righting without external rewards.
- Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies
-
Addressing the fundamental barrier that vision-language-action (VLA) policy actions cannot be interpreted via text vocabulary projections, this paper introduces an event-grounded sparse autoencoder (SAE) pipeline that anchors feature discovery in kinematic keyframe clusters and validates feature causality via closed-loop residual-preserving latent edits.
- Evo-Depth: A Lightweight Depth-Enhanced Vision-Language-Action Model
-
Addressing the spatial perception limitations of 2D visual representations in VLA models and the heavy deployment overhead of explicit 3D sensors or massive geometry models, Evo-Depth proposes a 0.9B lightweight framework that extracts multi-view implicit depth cues via IDEM and modulates vision-language representations via SEM, achieving state-of-the-art manipulation success rates with minimal GPU memory (3.2 GB) and high control frequency (12.3 Hz).
- Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models
-
This paper proposes Adaptive Failure-Informed Learning (AFIL), an end-to-end framework that trains Dual Action Generators sharing a VLM backbone on successful and online failure rollouts, adaptively repelling action sampling away from failure modes based on cosine score disagreement during denoising.
- Fast-WAM: Do World Action Models Need Test-Time Future Imagination?
-
To overcome the severe test-time latency caused by iterative future video denoising in World Action Models (WAMs), Fast-WAM disentangles video modeling during training from explicit future imagination during inference, executing direct action generation from single-pass world representations to achieve real-time 190ms control and high data efficiency without embodied pretraining.
- FastDSAC: Unlocking the Potential of Maximum Entropy RL in High-Dimensional Humanoid Control
-
Addressing the severe exploration collapse and critic overestimation in high-dimensional humanoid control, FastDSAC introduces Dimension-wise Entropy Modulation (DEM) for autonomous budget redistribution across actuators and a continuous Gaussian distributional critic, substantially surpassing deterministic SOTA baselines in massively parallel regimes.
- Feedback World Model Enables Precise Guidance of Diffusion Policy
-
Addressing the issue where open-loop world model prediction drift misguides diffusion policy denoising under distribution shift, this paper proposes an inference-time feedback world model that maintains a lightweight latent state to absorb real-time observation residuals and combines it with counterfactual variance-weighted action-aware guidance, substantially cutting prediction error and boosting policy success without retraining or extra data.
- FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies
-
FineVLA introduces an open framework for action-instruction alignment, constructing FineVLA-Data (47,159 human-verified fine-grained trajectories) and RoboFine-Bench, and demonstrating that a 1:1 mixture of fine-grained and raw instructions significantly enhances factor-level steerable control without sacrificing goal-level task success.
- FLASH: Efficient Visuomotor Policy via Sparse Sampling
-
FLASH parameterizes robot action trajectories into a compact orthogonal Legendre polynomial coefficient space, expands the physical control horizon via temporal sparse sampling, initiates single-step Flow Matching from an informative history polynomial prior, and directly provides analytic velocity feed-forward to dramatically reduce low-level torque tracking errors.
- FlowHOI: Flow-based Semantics-Grounded Generation of Hand-Object Interactions for Dexterous Robot Manipulation
-
FlowHOI introduces the first two-stage conditional flow-matching framework that decouples geometry-centric grasping from semantics-centric manipulation, combining an egocentric video reconstruction prior with 3D Gaussian scene features to generate physically plausible, semantically aligned HOI trajectories with a 40x speedup over diffusion baselines.
- Frequency-Aware Flow Matching for Continuous and Consistent Robotic Action Generation
-
Addressing identifiability failures under heterogeneous control frequencies and high-frequency jitter during inference caused by standard discrete action chunks, this paper introduces Frequency-Aware Flow Matching (FAFM), which performs flow matching over physical-time-anchored DCT coefficient space with analytic first-order temporal derivative regularization to generate continuous, temporally consistent robotic actions.
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control
-
Reframing generative robotic action synthesis from fixed-schedule time integration to iterative optimization over a time-invariant stationary velocity field (GeCO), enabling complexity-adaptive computation allocation and zero-overhead out-of-distribution anomaly detection.
- GeoWind2Plan: Mission-Time 3D Urban Wind Prediction for Energy-Efficient UAV Planning
-
GeoWind2Plan converts building geometry and background wind into a mission-corridor 3D mean wind field and jointly optimizes UAV path and speed; on block A at a background speed of 4 m/s, CFD-evaluated energy decreases from \(13.3\pm1.8\) for wind-agnostic planning to \(12.5\pm1.5\) Wh/km, while wind inference for a 20% corridor takes about 3 seconds.
- Grasp-Then-Plan with Failure Attribution: A Closed Two-Stage Framework for Precise and Generalizable Robotic Manipulation
-
To resolve the failure-attribution ambiguity caused by the tight coupling between grasping and motion planning, GTP-FA introduces a closed-loop two-stage framework that fuses a failure discriminator with a diagnostic representation space to direct optimization pressure toward task-aware grasp selection or planning policy refinement.
- HABIT: Human-Aware Behavior and Interaction Training Dataset for Robot Manipulation
-
To overcome the inability of robot policies trained in human-absent settings to coordinate safely with humans, this paper introduces HABIT, the first large-scale human-present bimanual demonstration dataset (10,563 episodes, 164 hours across 60 tasks) structured into Collaborator, Coworker, and Supervisor roles, demonstrating that reactive collection protocols elicit proactive yielding, gesture grounding, and spatiotemporal synchronization in downstream VLA policies.
- Hide-and-Seek in Trajectories: Discovering Failure Signals for VLA Runtime Monitoring
-
Hide-and-Seek formulates VLA failure detection as a coarsely supervised learning problem, utilizing dual-granularity inter- and intra-trajectory contrastive losses to isolate failure-indicative actions from trajectory-level supervision alone, delivering accurate, millisecond-level runtime monitoring and test-time policy steering via functional conformal prediction.
- HiDRA: Hierarchical Dual-Routing Attention for Replay-Free Lifelong Imitation Learning
-
HiDRA proposes a hierarchical dual-routing attention mechanism paired with key-space geometric regularization, decoupling instruction-level anchor retrieval from context-level dynamic refinement to overcome routing interference and eliminate catastrophic forgetting in replay-free, single-pass lifelong imitation learning.
- How to Instruct Your Robot: Dense Language Annotations Power Robot Policy Learning
-
DeMiAn leverages frozen VLMs to augment existing robot and human demonstration corpora across four complementary dense language aspects (physical motion, scene composition, arm pose, and reasoning), and deploys a lightweight asynchronous instructor to dynamically supply task-optimal annotations, substantially improving policy success rates without added execution latency.
- Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating
-
To tackle easy-reward bias and catastrophic forgetting in multi-object humanoid loco-manipulation across cluttered indoor scenes, Humanoid Horizon introduces a single-policy framework combining a parallel training strategy, dynamic starting, and recursive reward gating, achieving over 80% success on two-object benchmarks and graceful scaling to five consecutive transport cycles.
- HumanoidArena: Benchmarking Egocentric Hierarchical Whole-Body Learning
-
Introduces HumanoidArena, a simulation-first benchmark for egocentric hierarchical whole-body humanoid learning that unifies high-level visual decisions and low-level dynamic execution via a compact 40D canonical intermediate action interface across 7 leg-critical HOI/HSI tasks, revealing severe tracker conditioning and asymmetric cross-GMT transfer failures.
- ICAT: Incident-Case–Grounded Adaptive Testing for Physical-Risk Prediction in Embodied World Models
-
ICAT establishes an adaptive physical-risk evaluation benchmark for embodied world models grounded in authoritative incident reports and safety guidance, revealing that modern video-generative world models systematically omit or attenuate catastrophic consequences even after successfully executing hazardous actions.
- ICAT: Incident-Case-Grounded Adaptive Testing for Physical-Risk Prediction in Embodied AI
-
Addressing the safety illusion where video-generative embodied world models fail to anticipate hazardous physical outcomes, ICAT introduces an adaptive testing benchmark grounded in real incident reports and safety guidance, exposing that current generative simulators frequently omit severe consequences despite executing risky actions.
- Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator
-
Image2Sim decouples 3D spatial anchoring from photorealistic observation synthesis using feed-forward feature Gaussians and a geometry-aware one-step pixel flow renderer, driving an automated data engine that synthesizes over 10M navigation samples across 20K scenes to unlock zero-shot cross-simulator and real-world transfer.
- JODA: Composable Joint Dynamics for Articulated Objects
-
JODA presents a composable joint-level dynamics framework for articulated objects that models conservative forces, dry friction, and damping as a continuous three-channel field over the joint degree of freedom, enabling automated multimodal VLM generation, interactive editing, and gradient-based trajectory optimization.
- JRDB-AVR: An Active Visual Reasoning Benchmark for Embodied Agents in Real-World Environments
-
This paper introduces JRDB-AVR, the first real-world active visual reasoning benchmark that requires embodied agents to actively acquire observations across time and viewpoints while evaluating both final answers and localized bounding box evidence, and proposes JRDB-AVR-Agent, a training-free baseline using an explicit observation-grounded graph world model.
- LangMap: A Human-Verified Benchmark for Hierarchical Open-Vocabulary Goal Navigation
-
Introduces HieraNav and LangMap, an open-vocabulary embodied navigation benchmark with human-verified annotations across four hierarchical levels (scene, room, region, instance), together with PlaNaVid, a decoupled RGB-only baseline that outperforms heavy 3D-map methods without requiring depth, point clouds, or masks.
- Language-Conditioned World Modeling for Visual Navigation
-
The paper introduces a language-conditioned visual navigation dataset and compares two alternative approaches—“diffusion world model + latent-space actor–critic” and “unified autoregressive action/observation prediction”: the former has stronger image structural fidelity in seen environments, whereas the latter predicts offline trajectories better in unseen environments, without validating real closed-loop control.
- Learning from Trials and Errors: Reflective Test-Time Planning for Embodied LLMs
-
This paper introduces Reflective Test-Time Planning, which integrates pre-execution reflection-in-action (internal simulation and candidate selection) with post-execution retrospective reflection-on-action (test-time training for both action policy and internal reflection model), enabling embodied LLMs to continuously learn from deployment failures.
- Learning Human-Intention Priors from Large-Scale Human Demonstrations for Robotic Manipulation
-
MoT-HRA factorizes robot manipulation into 3D spatial trajectory prediction, MANO hand-motion intention modeling, and embodiment-specific action generation trained on a 2.2M-episode human demonstration dataset HA-2.2M with read-only KV transfer to prevent negative interference.
- Learning Visual Feature-Based World Models via Residual Latent Action
-
Addressing blurriness and collapse in direct feature regression alongside the curse of dimensionality in high-dimensional feature generation, this paper learns compact Residual Latent Actions (RLA) from DINO residuals and predicts them via flow matching, building an accurate, lightweight world model (RLA-WM) that enables actionless video imitation and visual RL entirely inside the world model.
- LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation
-
LIBERO-PeRM establishes a simulation benchmark for personalized robotic manipulation featuring a three-tier taxonomy of observable preferences (motion, skill, and user profile) across 312 tasks and 40.8k demonstrations, revealing fundamental limitations of current VLA policies in zero-shot personalization, out-of-distribution generalization, and conflict resolution.
- Linguistic Trajectory Encoding for Efficient Long-Horizon Spatial Memory in Embodied Agents
-
The paper encodes object motion histories as linked linguistic captions, sparse 3D anchors, and visual anchors, improving semantic trajectory retrieval in long egocentric videos while trading expensive one-time construction for sub-second online queries.
- Local Guidance, Global Impact: Gaussian-Reshaped Trust Region Unlocks Behavior Transitions
-
To resolve the failure of Proximal Policy Optimization (PPO) in continual and non-stationary reinforcement learning caused by directionally inefficient local exploration and monotonic divergence penalties hindering behavioral adaptation, this paper proposes Gaussian Trust Region Policy Optimization (GTR), which reshapes the trust region with a non-monotonic Gaussian kernel and incorporates mixture Gaussian anchors to unify local Fisher-geometric stiffness with seamless behavior mode transitions under strong advantages.
- M2E-UAV: A Benchmark and Analysis for Onboard Motion-on-Motion Event-Based Tiny UAV Detection
-
To tackle severe ego-motion background clutter and extreme target sparsity when both the observer and target UAV move simultaneously, M2E-UAV establishes the first onboard motion-on-motion event-based benchmark with synchronized 720p event-IMU streams and point-level foreground supervision, diagnosing systematic bottlenecks across event-frame, voxel-grid, and point-set detectors.
- ManiTaskGen: A Comprehensive Task Generator for Benchmarking and Improving Vision-Language Agents on Embodied Decision-Making
-
ManiTaskGen addresses the task scarcity and scene dependency of embodied decision-making by grounding task synthesis in the agent-environment interaction cycle and receptacle-aware 3D scene graphs, automatically generating logically comprehensive process- and outcome-based mobile manipulation tasks for arbitrary scenes while enabling inference-time self-reflection optimization.
- Masked Visual Actions for Unified World Modeling
-
This paper proposes Masked Visual Actions, which casts action as partially revealed spatiotemporal entity trajectories directly in pixel space, enabling a single video foundation model finetuned on just 15 hours of data to unify forward physical simulation and inverse action extraction in a zero-shot manner.
- METIS: Multi-Source Egocentric Training for Integrated Dexterous Vision-Language-Action Model
-
METIS builds EgoAtlas, a multi-source egocentric dexterous manipulation dataset comprising 343K trajectories, introduces a compact motion-aware dynamics representation (4 visual tokens + 40 motion tokens), and unifies chain-of-thought reasoning and low-level action generation via adaptive token gating in a single VLA model, achieving the highest average success rate on real-world dual dexterous humanoid robots.
- MineEvolve: Self-Evolution with Accumulated Knowledge for Long-Horizon Embodied Minecraft Agents
-
MineEvolve introduces a training-free self-evolution framework for embodied agents that converts execution traces into typed feedback, induces reusable skills and failure remedies, and performs prefix-frozen local plan repair on long-horizon tasks.
- MineEvolve: Self-Evolution with Accumulated Knowledge for Long-Horizon Embodied Tasks
-
MineEvolve establishes a parameter-free self-evolution framework for embodied agents that converts typed execution feedback into reusable skills and localized failure remedies, enabling budget-constrained local plan repair and achieving major performance gains across complex long-horizon Minecraft tasks.
- MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction
-
MolmoMotion formalizes the task of forecasting 3D world-coordinate trajectories of object surface points conditioned on language instructions, scales up supervision with a 1.16M unconstrained video dataset alongside PointMotionBench, and demonstrates significant improvements over existing motion baselines with seamless transfer to robot manipulation and video generation.
- Nautilus: From One Prompt to Plug-and-Play Robot Learning
-
Nautilus proposes a domain-specific agent harness for robot learning research that leverages typed interface contracts, chambered execution, and a uniform two-tier transport substrate to turn natural language prompts into reproducible, plug-and-play evaluation and deployment workflows across diverse policy families, benchmarks, and real robots.
- NDPP-Grasp: Non-Differentiable Physical Plausibility Constraint-Guided Task-Oriented Dexterous Grasp Generation
-
Addressing the non-differentiable nature of physical constraints (such as mesh penetration) in task-oriented dexterous grasp synthesis, NDPP-Grasp introduces a gradient-free stochastic optimal control law coupled with an amortized lookahead strategy into the diffusion denoising trajectory, delivering real-time execution alongside high physical plausibility and task alignment.
- NDPP-Grasp: Non-Differentiable Physical Plausibility Constraint-Guided Task-Oriented Grasp Generation
-
Addressing the non-differentiable nature of physical plausibility constraints (non-penetration, self-collision, and joint limits) in task-oriented dexterous grasping, NDPP-Grasp formulates denoising as stochastic optimal control to derive a gradient-free guidance law paired with an amortized lookahead strategy, injecting physical constraints progressively throughout denoising with 11–20 ms real-time inference.
- NestedVLA: Learning to Consolidate and Generate Skills for Vision-Language-Action Models
-
NestedVLA tackles parameter explosion and catastrophic weight interference in post-hoc skill consolidation for vision-language-action (VLA) models by shifting from parameter aggregation to parameter generation: using an embedded patch-based hypernetwork to synthesize action-expert deltas while retaining separate compressed VLM residuals, it consolidates three diverse manipulation policies at 1.5x checkpoint storage and reaches 66.7% average success (outperforming the strongest dynamic merging baseline by 41.1 percentage points).
- neuro-inspired inverse learning for planning and control
-
This paper introduces an Inverter framework inspired by mammalian motor control that amortizes multi-step trajectory planning by directly differentiating Bolza control objectives through a frozen forward dynamics model, matching or beating offline RL and diffusion planners across D4RL while delivering 1-to-2 orders of magnitude faster inference.
- Next Forcing: Causal World Modeling with Multi-Chunk Prediction
-
Addressing the myopic supervision and appearance shortcut in autoregressive world action models caused by extreme visual similarity between consecutive frames, Next Forcing introduces a multi-chunk prediction framework with chained auxiliary modules and multi-layer feature fusion, accelerating training convergence, elevating physical fidelity, and delivering a 2× inference speedup.
- Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models
-
Through closed-loop causal interventions, sparse autoencoders, and physical robot experiments across six VLA architectures over 420,000+ rollouts, this work provides the first mechanistic proof that fine-tuned VLA policies execute coordinate-bound motor programs anchored to absolute workspace locations rather than tracking objects, while establishing a clear functional division between VLM goal selection and motor expert execution.
- NoTVLA: Semantics-Preserving Robot Adaptation via Narrative Action Interfaces
-
NoTVLA introduces a semantics-preserving robot adaptation interface that replaces dense low-level action imitation with sparse narrative waypoints, anchor-conditioned depth grounding, and deterministic B-spline detokenization, effectively halting catastrophic forgetting while dramatically improving semantic out-of-distribution generalization and camera robustness.
- OASIS: Observation-Action Space Alignment via SE(3) Trajectory Prediction for Robotic Manipulation
-
Addressing the geometric misalignment between 2D observation spaces and rigid-body action manifolds in current visuomotor policies, OASIS predicts camera-frame \(SE(3)\) end-effector trajectories as intermediate latent representations, allowing a lightweight action decoder to achieve state-of-the-art manipulation across LIBERO, CALVIN, and real-world robots with high sample efficiency.
- ODEWorld: A Continuous Predictive Architecture via Physical-Time Flow
-
Presents ODEWorld, a continuous-time latent world model built upon Physical-Time Flow (PT-Flow), which explicitly decouples static context from dynamics and applies direct first-order velocity supervision via Jacobian-vector products, reformulating future state prediction into compact ODE integration while eliminating representation collapse.
- OGPO: Offline Goal-conditioned Policy Optimization for Recoverable Vision-Language-Action Models
-
OGPO introduces an offline goal-conditioned policy optimization framework for VLA post-training that adaptively relabels demonstrations into key-state segments with process rewards and AWAC-style advantage updates, dramatically enhancing long-horizon manipulation and execution-time recoverability without online rollouts.
- OmniDex: Scaling Dexterous Hand Grasping to Diverse Cluttered Scenes
-
Addressing the critical scarcity of cluttered dexterous grasping data and the "last-millimeter" precision gap in multi-contact synthesis, OmniDex constructs the first benchmark spanning 2.6 million scenes and 400 million physically verified grasps via a decoupled seed-and-filter protocol, paired with an end-to-end physics-constrained diffusion framework with Soft-WTA multimodal learning that enables fast, test-time optimization-free execution.
- One Scan is Enough: Demonstration-Free Adaptation for Language-Guided Navigation
-
Addressing the deployment failure of language-guided navigation VLAs in unfamiliar environments, this paper introduces a framework that transforms a single handheld scan into compiled training episodes via Scan2Instruction and adapts the policy via PRPT-KS using critic-free closed-form Pareto geometric utilities, achieving an 85% real-robot success rate without human demonstrations.
- OntoPlan: An Ontology-Grounded Scene Representation and Agentic Framework for Scalable Robot Task Planning
-
To tackle spatial context hallucination, precondition violations, and token explosion in LLM-based long-horizon robot task planning in large environments, OntoPlan proposes an ontology-grounded world model strictly aligned with PDDL and a collaborative agent framework, achieving 0.89 task success and 5.6× lower token usage over baselines.
- PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking
-
PACE introduces a training-free test-time execution method that identifies manipulation phase transitions from low-speed valleys in the predicted action chunk's kinematic profile, dynamically selecting execution horizons to consistently surpass fixed-horizon baselines on RoboTwin2.0 and real ALOHA/Franka robots.
- PDF-HR: Pose Distance Fields for Humanoid Robots
-
Introduces Pose Distance Fields for Humanoid Robots (PDF-HR), a continuous and differentiable implicit pose manifold prior that maps high-dimensional joint configurations to geodesic distances, serving as a plug-and-play term to accelerate RL motion tracking and regularize motion retargeting.
- PhysGuard: Fisher-Guided Gradient Projection for Sim-to-Real Neural PDE Surrogates
-
To prevent neural PDE surrogates from destroying pretrained large-scale physical priors during sim-to-real adaptation, PhysGuard leverages a kernel Gram-matrix formulation of the empirical Fisher Information Matrix to identify physics-critical parameter directions, projecting fine-tuning gradients onto their orthogonal null space to reduce low-frequency error by up to 32%.
- PO-PDDL: Learning Symbolic POMDPs from Visual Demonstrations for Robot Planning Under Uncertainty
-
PO-PDDL introduces an LLM-friendly symbolic formulation unifying PDDL syntax and POMDP uncertainty, learning stochastic transitions and observation models from video demonstrations via latent trajectory reconstruction and visual contradiction discovery for robust, low-latency belief-space robot planning.
- Point Tracking Improves World Action Models
-
This paper introduces JOPAT, a JOint Pixel-And-Track world-action model that jointly denoises visual latents, 2D point tracks with visibility, and robot actions within a single diffusion transformer, establishing explicit motion correspondences as the future-state interface for control and achieving a state-of-the-art 97.8% success rate across 40 LIBERO tasks.
- Pose6DAug: Physically-Plausible Multi-View Object Swapping for Robot Data Augmentation
-
Addressing the challenge of vision-language-action (VLA) policy failure on novel objects and the high cost of multi-view teleoperation, Pose6DAug introduces a failure-driven 3D data augmentation framework that swaps manipulated objects in successful episodes using continuous 6D pose trajectories and 3D meshes with physically grounded geometric perturbations, generating strictly multi-view consistent and kinematically valid demonstrations without collecting new data and boosting out-of-distribution success rate by 16.5% over the prior state-of-the-art.
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models
-
PriorVLA introduces a prior-preserving adaptation framework that decouples prior preservation from task specialization using Dual Action Experts (a frozen Prior Expert and a trainable Adaptation Expert) and Expert Queries to capture scene and motor priors, updating only 25% of full fine-tuning parameters while dramatically boosting OOD generalization and few-shot manipulation.
- ProcObject-10K: Benchmarking Object-Centric Procedural Understanding in Instructional Videos
-
ProcObject-10K establishes a benchmark of 10,522 grounded QA pairs across 1,799 instructional video clips to evaluate object-centric procedural understanding across dual viewpoints, exposing a substantial answering-grounding gap (mIoU << 45%) in top MLLMs and providing a spatial-temporal regularized SFT framework that transfers to embodied planning.
- ProCompNav: Proactive Instance Navigation with Comparative Judgment for Ambiguous User Queries
-
ProCompNav collects same-category objects in an unknown environment, then recursively identifies distinguishing attributes and asks users yes/no questions, achieving success rates of 23.7%, 28.1%, and 17.0% across the three simulated CoIN-Bench splits while substantially shortening user-simulator responses.
- Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs
-
To overcome the severe latency bottleneck of full iterative denoising in diffusion-based vision-language-action models during robot replanning, Realtime-VLA FLASH introduces a dual-path speculative inference framework—combining a 1-block Gemma draft model, parallel endpoint verification via flow-matching interpolation paths, and a gripper-triggered phase-aware fallback—achieving a 3.04× speedup on LIBERO and enabling reactive real-world conveyor-belt sorting up to 15 m/min with negligible loss in task success.
- RedVLA: Physical Red Teaming for Vision-Language-Action Models
-
Addressing the unpredictable and irreversible physical harm in Vision-Language-Action (VLA) deployment, this paper proposes RedVLA—the first physical red teaming framework that injects minimal risk factors into critical interaction regions and refines them via trajectory-guided zero-order optimization, achieving up to 95.5% attack success rate while maintaining task feasibility and deriving a lightweight safety guardrail SimpleVLA-Guard.
- RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models
-
RePO-VLA introduces a recovery-driven policy optimization framework that eliminates causal confusion via history-reset trajectory slicing, assigns dense progress labels via a self-referential semantic value function, and conditions policy inference on a constant high value (\(v=1.0\)) to enable autonomous recovery onto the success manifold without external failure detectors.
- RescueBench: Can Embodied Agents Save Lives in the Wild?
-
This paper introduces RescueBench, a photo-realistic diagnostic benchmark for embodied search-and-rescue (SAR) that models missions as a four-stage sequential pipeline (exploration, rescue, memory-guided return, and handoff) across five progressive difficulty levels, revealing that zero existing embodied architectures complete the full pipeline at the highest difficulty due to persistent exploration and spatial memory bottlenecks.
- Resilience Matters for Embodied Agents System: New Metrics, Systematic Evaluation, and Optimization
-
Addressing the issue where outcome-centric metrics hide dynamic execution pathologies in embodied agents, this paper adapts resilience engineering principles to establish an evaluation layer covering rebound, stability, and graceful extensibility, enabling targeted non-intrusive execution optimization.
- Rethinking Causal Action Tokenization with Conditional Annealing in Flow Matching
-
CATok reframes continuous action tokenization as a conditional annealing flow-matching process, endowing discrete action tokens with a coarse-to-fine causal hierarchy that substantially boosts both training efficiency and manipulation performance in purely autoregressive VLAs.
- Reverse to Advance: Teleoperation-Cost Effective Hard Policy Learning from Reversed Easy Tasks
-
Addressing the bottleneck of prohibitive teleoperation costs in high-precision contact manipulation, Auto-E2H exploits directional task difficulty asymmetry by combining closed-loop autonomous resetting, hierarchical kinematic-and-critic refinement, and iterative diffusion policy learning to achieve high success rates with minimal human demonstration effort.
- Revisiting Embodied Chain-of-Thought for Generalizable Robot Manipulation
-
This paper revisits embodied chain-of-thought in robotic manipulation by revealing the scaling bottlenecks of autoregressive reasoning prefixes and proposing ERVLA—a framework integrating an auxiliary choice policy, knowledge-truncated KV conditioning, and reasoning dropout to internalize reasoning into continuous flow-matching policies.
- RoboWits: Unexpected Challenges for Robotic Creative Problem Solving
-
Introduces RoboWits, the first robotic benchmark dedicated to evaluating bi-manual creative problem-solving and strategy adaptation under unexpected physical challenges, powered by an automated multi-agent task generation and mutation pipeline that systematically exposes the reasoning brittleness of modern VLAs.
- ROCKET: Residual-Oriented Multi-Layer Alignment for Spatially-Aware Vision-Language-Action Models
-
ROCKET introduces a residual-oriented multi-layer representation alignment framework that eliminates cross-layer gradient interference via a shared projector and balances depth-wise supervision using Matryoshka-style sparse activation, achieving a 98.5% state-of-the-art success rate on LIBERO with only ~4% of previous training compute.
- RotVLA: Rotational Latent Action for Vision-Language-Action Model
-
RotVLA introduces a continuous \(\mathrm{SO}(n)\) rotational latent action representation trained with a triplet temporal compositionality objective, which acts as a high-level latent planner within a unified flow-matching action expert to achieve state-of-the-art results on LIBERO, RoboTwin2.0, and real-world manipulation with only 1.7B parameters.
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
-
Under a strict offline, model-free robotic setting with only fragmented, task-agnostic trajectories, SAGAS decouples graph-assisted trajectory stitching into an offline reusable reachability interface and test-time semantic graph augmentation, synthesizing cost-aware accepting prefix–suffix plans via Büchi product search for zero-shot LTL execution.
- Self-Supervised Doppler-Guided RF Odometry
-
The paper introduces RFWalk, a self-supervised, RF-only odometry framework that models cross-frame dense correspondences between range-azimuth heatmaps as random walks on a space-time graph and leverages radar-measured Doppler radial velocities as local first-order physical constraints, reducing absolute trajectory error by 58% in dynamic pedestrian scenes without ground-truth poses or auxiliary sensors.
- SENTINEL: A Multi-Level Formal Framework for Safety Evaluation of Foundation Model-based Embodied Agents
-
SENTINEL grounds embodied agent physical safety requirements in formal temporal logic specifications (LTL/CTL) across semantic interpretation, high-level planning, and branching trajectory execution, uncovering failure modes that subjective LLM evaluators miss and enabling verifiable counterexample-driven safety refinement.
- Shared Modular Recurrence in Contextual MDPs for Universal Morphology Control
-
This paper frames multi-robot morphology control as a Contextual MDP with partially observable context, introducing shared modular recurrence across robot limbs into transformer architectures to perform implicit system identification from local histories, substantially improving zero-shot generalization to unseen dynamics, kinematics, and topologies.
- Simulation-Informed Diffusion for Decentralized Multi-robot Motion Planning
-
To overcome the myopia of local observations and the unreliability of communication in decentralized multi-robot motion planning (MRMP), this paper proposes Simulation-Informed Diffusion (SID), which uses a single constraint-aware diffusion model both to simulate dynamically feasible neighbor trajectories and to plan the robot's own collision-free path, paired with an on-demand minimal communication scheme that resolves symmetry deadlocks while keeping communication near zero.
- SkiP: When to Skip and When to Refine for Efficient Robot Manipulation
-
SkiP reformulates temporal abstraction in robot manipulation through a training-target relabeling mechanism and unsupervised Motion Spectrum Keying (MSK), replacing behavior cloning targets in free-space skip segments with the entrance action of the next key segment to achieve adaptive skipping and refinement within a single policy, cutting executed steps by 15%–40% while improving task success across 72 simulated and 3 real-robot tasks.
- SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework
-
Addressing the severe occlusion, contact artifacts, and poor generalization of raw video-derived kinematic priors, SUGAR establishes an end-to-end learning framework featuring automated prior extraction, privileged physics refinement, and hierarchical autonomous distillation, achieving zero-shot deployment and closed-loop failure recovery across six loco-manipulation tasks on a Unitree G1 humanoid.
- SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators
-
Addressing the core barrier where numerical control commands induce conflicting pixel motions across diverse camera views and robot setups, SyncWorld introduces a 6-DoF in-context visual calibration episode alongside calibration distillation, enabling zero-shot simulation and training-free test-time policy improvement across unseen setups.
- Tactile MNIST: Benchmarking Active Tactile Perception
-
Introduces the first open-source, Gymnasium-compatible benchmark suite for active tactile perception (TMBS and ap_gym), accompanied by 13,580 high-resolution 3D-printable digit models, 153,600 real-world robotic GelSight Mini contacts, and a standardized dual-efficiency evaluation protocol.
- TGPO: Temporal Grounded Policy Optimization for Signal Temporal Logic Tasks
-
TGPO proposes a hierarchical reinforcement learning framework that decomposes Signal Temporal Logic (STL) specifications into sequenced subgoals with invariant constraints, grounds temporal variables via critic-guided Metropolis-Hastings Bayesian sampling, and trains low-level time-conditioned policies with dense stage-wise rewards, outperforming existing RL baselines by an average of 31.6% in success rate across diverse high-dimensional robotic tasks.
- TGPO: Trace-Guided Policy Optimization for Robot Task Planning via Verifiable Subgoal Generation
-
Addressing the core tension where LLMs produce infeasible plans while symbolic planners struggle with abstract human instructions, TGPO formulates task planning as verifiable PDDL subgoal generation and introduces a reinforcement learning framework combining verifier-grounded rewards, intermediate trace correction, and constrained policy updates to achieve robust long-horizon planning.
- Think Densely, Act Sparsely: Latent Expert Cognitive Chains for Vision-Language-Action Autonomous Driving
-
To resolve the critical scene information loss caused by sparse trajectory supervision in Vision-Language-Action (VLA) models, this paper introduces the "Think Densely, Act Sparsely" paradigm and the LECDrive framework, which embeds a progressive representation-semantics-geometry latent cognitive chain and a hybrid motion primitive refinement scheme, achieving new state-of-the-art results across closed-loop planning and driving perception benchmarks without incurring inference overhead.
- Trajectory Planning without Trajectory Data: A Manifold-Guided Approach
-
Addressing the core limitations of generative trajectory planners—namely their reliance on expensive expert demonstrations and severe vulnerability when generalizing to unseen start-goal pairs—this paper presents Ariadne, a framework that extracts state-space manifold geometry from state-only observations and integrates a Correction-Projection ODE to synthesize feasible paths without any trajectory supervision.
- UniSteer: Unified Noise Steering for Efficient Human-Guided VLA Adaptation
-
Addressing the core tension where flow-matching VLA adaptation suffers from inefficient autonomous exploration while human interventions cannot directly supervise frozen noise-space policies, UniSteer introduces a contraction-mapping fixed-point inversion to recover high-fidelity noise targets from human actions, unifying imitation and online RL under a shared noise interface to raise average real-world success rates from 20% to 90% in 66 minutes.
- UniT: Toward a Unified Physical Language for Human-to-Humanoid Policy Learning and World Modeling
-
UniT introduces a visually anchored, tri-branch cross-reconstruction latent action tokenizer that projects heterogeneous human and humanoid movements into a shared discrete physical intent space, powering state-of-the-art VLA policy generalization, zero-shot task transfer, and cross-embodiment world model control.
- URDF-Anything+: End-to-End Generation for Simulation-Ready Articulated Assets
-
URDF-Anything+ introduces an end-to-end autoregressive diffusion framework that directly generates simulation-ready URDF models with rigid link meshes and kinematic joint parameters from a single RGB image, enabling zero-shot real-robot policy transfer under a Real-Follow-Sim paradigm.
- VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation
-
VersaCamVLA establishes a camera-configurable Vision-Language-Action framework that decouples camera-set representation from policy action learning, leveraging natural wrist-camera motion for zero-cost pose augmentation and multi-signal target-view prediction to map arbitrary posed views into fixed-size scene tokens, achieving robust manipulation across varying camera counts and unseen camera poses without explicit 3D sensing.
- VLA-Pro: Cross-Task Procedural Memory Transfer for Vision-Language-Action Models
-
To tackle cross-task generalization bottlenecks and execution biases in vision-language-action models, VLA-Pro constructs a retrievable procedural memory bank of structured states and task-specific LoRA adapters, dynamically retrieving and fusing adapters at runtime for plug-and-play experience transfer to unseen tasks.
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models
-
Addressing the high hardware cost and poor cross-laboratory reproducibility of real-world Vision-Language-Action (VLA) evaluations, VLA-REPLICA introduces a standardized ~$1,050 benchmark featuring controlled light-box illumination, dual AprilTag and video overlay calibration, universal action normalization, and 90 test scenes covering in-distribution adaptation and out-of-distribution generalization, enabling highly reproducible decentralized local robot evaluations.
- VolCo: Volumetric Contact for High-Fidelity Human Grasp Generation
-
To overcome severe penetrations and loss of local detail inherent in surface-based contact representations, this paper presents VolCo, a volumetric contact representation that extends contact into 3D voxel grids, and VolCoDiff, a hierarchical latent diffusion model that achieves state-of-the-art stability and low penetration depth in human grasp synthesis.
- When and What to Prune? Stage-Aware Visual Token Pruning for Efficient VLA
-
SAPrune is a training-free visual token pruning framework for Vision-Language-Action (VLA) models that uses offline attention dynamics to identify stable redistribution plateaus for layer selection, combined with an attention-core and functional-diversity dual-path retention mechanism, pruning 87.5% of visual tokens (down to 32) and achieving up to 1.718× inference speedup with negligible success rate loss.
- When Simulation Lies: A Sim-to-Real Benchmark and Domain-Randomized RL Recipe for Tool-Use Agents
-
Formalizing the deployment brittleness of tool-use language agents as a partially observable Markov decision process (POMDP) sim-to-real gap, this work introduces RobustBench-TC with 22 production-grounded perturbation types, and demonstrates that domain-randomized reinforcement learning (ToolRL-DR) on static perturbations zero-shot transfers to close roughly 27% of unseen runtime transition failures.