Skip to content

๐Ÿค– Robotics & Embodied AI

๐Ÿ’ฌ EMNLP2026 ยท 7 paper notes

๐Ÿ“Œ Same area in other venues: ๐Ÿง  NeurIPS2026 (131) ยท ๐ŸŽž๏ธ ECCV2026 (135) ยท ๐Ÿ“ท CVPR2026 (163) ยท ๐Ÿ”ฌ ICLR2026 (166) ยท ๐Ÿ’ฌ ACL2026 (11) ยท ๐Ÿงช ICML2026 (54)

๐Ÿ”ฅ Top topics: Robotics ร—2

Inferring the Unspoken: Aligning Embodied Agents with Implicit Preferences

Addressing the failure of embodied agents to satisfy idiosyncratic user preferences under underspecified natural language commands, this paper introduces the 290-preference PBP benchmark and the InTU framework, demonstrating that verbalizing latent preferences prior to planning cuts alignment error roughly in half and localizes the bottleneck to preference acquisition rather than downstream execution.

IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation

Addressing the issue where current-frame observation aliasing under partial observability causes chunked policies to resample conflicting intents across adjacent steps, IntentVLA extracts compact camera and register tokens from recent visual history via a frozen VGGT encoder, injects them into the Qwen3-VL backbone via gated cross-attention along with a global intent summary token, and conditions a DiT flow-matching action head, substantially improving rollout stability and task success across AliasBench, SimplerEnv, LIBERO, and real-world manipulation.

LIBERO-Para: A Diagnostic Benchmark and Metrics for Paraphrase Robustness in VLA Models

This paper introduces LIBERO-Para, the first controlled diagnostic benchmark that independently decouples action and object expressions to evaluate paraphrase robustness in Vision-Language-Action (VLA) models, alongside PRIDE, a difficulty-grounded deviation metric. Across seven diverse VLA configurations, the study exposes a pervasive 22โ€“52 percentage point performance drop, demonstrates that object lexical variation acts as the primary bottleneck, and proves via DTW trajectory analysis that 80โ€“96% of failures stem from planning-level task identification divergence rather than motor control execution.

LookStep: Efficient Vision-Language Navigation with Linguistic Foresight and Event Driven Memory

Addressing the low data efficiency of next-step action imitation and the memory explosion from accumulating visual histories in continuous vision-language navigation, LookStep introduces linguistic foresight for candidate action consequences and an event-driven rolling episodic memory, achieving a 49.7% success rate on R2R-CE using only monocular RGB while cutting peak GPU memory by over 55%.

MineExplorer: Evaluating Open-World Exploration of MLLM Agents in Minecraft

Addressing the entanglement of game-specific mechanics and short-horizon objectives in existing benchmarks, MineExplorer establishes a knowledge-controlled open-world exploration benchmark in Minecraft with ReAct-style implicit multi-hop task graphs and multi-agent scene synthesis, revealing severe degradation in frontier MLLMs when coordinating hidden prerequisites.

Mini-BEHAVIOR-Gran: Revealing U-Shaped Effects of Instruction Granularity on Language-Guided Embodied Agents

Addressing the lack of cross-task quantification and controlled evaluation for instruction granularity in language-guided embodied AI, this paper introduces MINI-BEHAVIOR-GRAN, formalizes planning-width as a robust cross-task metric, and reveals a non-monotonic U-shaped performance pattern across VLA architectures driven by vision-dominant shallow grounding under coarse guidance.

when search becomes memory accelerating robot design discovery with self evolvin

Addressing the memoryless limitation of traditional robot morphology search where costly simulation traces are discarded after each iteration, this paper introduces AUTO-ROBOTIST, a self-evolving LLM agent that distills evaluation traces into an explicit natural-language skill library (featuring archetypes, positive/negative rules, and empirical observations), accelerating cold-start discovery and enabling effective cross-scale transfer from \(5\times 5\) to \(10\times 10\) design spaces.