AtomicVLA: Unlocking the Potential of Atomic Skill Learning in Robots¶

Conference: CVPR 2026
arXiv: 2603.07648
Institutions: Sun Yat-sen University, Peng Cheng Laboratory, Yinwang Intelligent Area: Robotic Manipulation / Vision-Language-Action Models
Keywords: VLA, Atomic Skills, Mixture-of-Experts, Continual Learning, Task Planning, Skill Routing

TL;DR¶

This paper proposes AtomicVLA, a unified planning-execution framework built upon \(\pi_0\). By utilizing adaptive Think-Act switching to generate atomic skill abstractions and routing actions to specialized experts via Skill-Guided MoE (SG-MoE), it improves the LIBERO-LONG success rate from 85.2% to 95.2% (+10%), achieves a +18.3% Gain in real-world Franka long-horizon tasks, and increases continual learning performance by 21%.

Background & Motivation¶

Limitations of Prior Work: Current VLA models (e.g., \(\pi_0\), OpenVLA) employ a single action decoder, mixing all skill knowledge within the same parameter set. They lack explicit planning capabilities for multi-step tasks and suffer from catastrophic forgetting when learning new skills.
Background: Methods like SayCan and Inner Monologue use external LLMs for high-level planning combined with independent low-level controllers. However, the lack of mutual awareness between the planner and executor leads to outdated or irrelevant instructions and system latency.
Goal: Robotic models must simultaneously support (1) high-level reasoning and task planning, (2) fine-grained action generation, and (3) scalable continual learning—capabilities that existing methods fail to provide concurrently.
Key Idea: To decompose complex tasks into reusable atomic skills (e.g., grasp, push, rotate), where each skill is handled by a dedicated expert. This modular design enables the continuous expansion of a skill library.

Method¶

Overall Architecture¶

AtomicVLA is an end-to-end framework based on \(\pi_0\) that unifies thinking and acting modalities. Given multi-camera observations \(O_t^{1:n}\) and a language instruction \(\ell\), the model adaptively predicts whether to enter thinking or acting mode:

Thinking Mode: Activated at task onset or sub-skill transitions. It generates three outputs: a task chain \(C_{0 \to k}\) (decomposing instructions into ordered sub-goals), current progress \(C_t\), and an atomic skill abstraction \(\sigma\).
Acting Mode: Activated during execution. Based on the latest skill abstraction \(\sigma\) and proprioceptive state \(s_t\), it generates action chunks \(A_t\) via SG-MoE.

Offline, kinematic data is used to generate supervision for "task chain + progress + skill labels." During runtime, the model switches adaptively between think and act modes. In the act phase, SG-MoE routes actions to specialized experts. If an anomaly or sub-skill transition is detected, the model returns to the think phase for re-planning. New skills are added to the library by extending the architecture without modifying existing components.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    subgraph DATA["Kinematics-based Embodied Reasoning Data Generation (Offline)"]
        direction TB
        D1["Demo Trajectories → Principal Axis Motion Analysis<br/>Segment Atomic Skill Snippets"] --> D2["InternVideo2.5 Semantic Correction<br/>→ Task Chain + Progress + Skill Labels"]
    end
    DATA -.Training Supervision.-> SW

    O["Multi-camera Observations O + Language Instruction ℓ"] --> SW{"Adaptive Think-Act Switching<br/>Predict [think] / [act]"}
    SW -->|think| TH["Thinking: Output Task Chain C₀→ₖ<br/>Progress Cₜ, Skill Abstraction σ"]
    SW -->|act| MOE["SG-MoE<br/>σ → Noise Level Zσ → Skill Router → top-1 expert"]
    TH --> MOE
    MOE --> EXE["Shared Expert + Skill Expert<br/>Weighted Action Chunk Generation"]
    EXE -->|Anomaly / Skill Switch| SW
    EXE -->|Task Completion| OUT["Long-horizon Task Completed"]

    CL["Modular Continual Learning: New Skills"] -.Freeze Old Experts · Train New Expert + Router Branch.-> MOE

Key Designs¶

1. Adaptive Think-Act Switching: Self-determined Planning vs. Execution

Running high-level planning at every frame is wasteful, as robots mostly execute pre-determined sub-goals. However, removing planning entirely causes models to lose track in multi-step tasks. AtomicVLA introduces [think] and [act] tokens into the vocabulary, allowing the model to decide its mode at each step. Planning occurs only at task starts or transition points, saving computation while enabling error recovery; if an object is dropped, the model detects the anomaly, switches to [think], and regenerates the skill abstraction.

2. SG-MoE: Semantic Skill Routing for Expert Specialization

Standard MoE routes at the token level, forcing experts to learn a mix of skills. SG-MoE uses the atomic skill abstraction \(\sigma\) as the routing signal. Each skill maps to a scalar noise level \(\sigma \in [0,100]\), converted into a high-dimensional vector \(Z_\sigma = E(\text{norm}(\log(\sigma)))\). The Skill Router calculates the distribution over \(Z_\sigma\) to activate the top-1 skill expert. The final action is a weighted combination of a shared expert (preserving general capability) and a specialized skill expert:

\[F_{out} = (1-w_k) \cdot F_{share}(x_t) + w_k \cdot F_k(x_t)\]

3. Modular Continual Learning: Scalable Expansion without Forgetting

Traditional VLA fine-tuning causes catastrophic forgetting (e.g., \(\pi_{0.5}\) loses 15% performance on old skills after learning new ones). AtomicVLA treats new skills as additions: when a new atomic skill is introduced, only a new randomly initialized expert is added, and the router's embedding space is expanded. Existing experts are frozen, ensuring zero degradation of prior knowledge without requiring EWC or experience replay.

4. Kinematic-based Embodied Reasoning Data Generation: Physical Priors over VLM Segmentation

Data for the thinking mode is generated using a two-stage pipeline. First, principal axis motion analysis of trajectories (tracking translation, rotation, and gripper states) automatically segments trajectories into atomic snippets (e.g., Z-axis descent + gripper closure = "pick"). Second, InternVideo2.5 provides semantic labels to refine these segments. This kinematic boundary detection is more precise and requires less human intervention than pure VLM-based methods.

Key Experimental Results¶

Main Results: LIBERO Benchmark (Success Rate %)¶

Method	Spatial	Object	Goal	Long	Avg.
Octo	78.9	85.7	84.6	51.1	75.1
OpenVLA	84.9	88.4	79.2	53.7	76.5
CoT-VLA	87.5	91.6	87.6	69.0	81.1
\(\pi_0\)	96.4	98.8	95.8	85.2	94.2
\(\pi_{0.5}\)	98.8	98.2	98.0	92.4	96.9
AtomicVLA	96.8	98.0	96.4	95.2	96.6
AtomicVLA*	98.8	98.8	97.2	96.2	97.8

Main Results: CALVIN ABC→D (Task Sequence Completion Rate %)¶

Method	1 Task	2 Tasks	3 Tasks	4 Tasks	5 Tasks	Avg.Len↑
\(\pi_0\)	94.3	87.0	77.9	68.5	59.4	3.87
\(\pi_{0.5}\)	91.9	84.6	79.4	75.5	71.0	4.02
AtomicVLA	95.0	87.8	81.9	75.0	69.1	4.09
AtomicVLA*	94.1	88.7	85.2	81.7	77.6	4.27

Real Franka Robot: Long-horizon Tasks (Success Rate %)¶

Method	Put in Plate	Put in Drawer	Put in Microwave	Avg.	ΔAvg.
\(\pi_0\)	45	55	10	36.7	—
\(\pi_{0.5}\)	65	35	35	45.0	—
AtomicVLA	65	60	45	56.7	+20.0
AtomicVLA*	75	60	55	63.3	+18.3

Ablation Study: SG-MoE Routing Mechanism (LIBERO-LONG %)¶

Method	LIBERO-LONG
\(\pi_0\) (No MoE)	85.2
+ Standard Token-level MoE	88.6 (+3.4)
+ MoDE (Denoising-step Routing)	89.5 (+4.3)
+ SG-MoE (Atomic Skill Routing)	95.2 (+10.0)

Key Findings¶

Significant Gains in LIBERO-LONG: AtomicVLA improves success rates by 10% on the most challenging long-sequence tasks, proving the value of explicit planning and skill decomposition.
Superiority of SG-MoE: Compared to standard token-level MoE (+3.4%) and MoDE (+4.3%), SG-MoE (+10.0%) ensures all tokens of a specific skill are processed by the same expert, minimizing cross-skill interference.
Minimal Forgetting: While \(\pi_{0.5}\) experiences a 15% drop in prior skill performance after learning new tasks, AtomicVLA* shows only a 1.3% decrease.
Error Recovery: The model demonstrates the ability to detect task failures and re-plan by generating new skill abstractions, though standard benchmarks like CALVIN may not fully capture this benefit.

Highlights & Insights¶

Unified Think-Act Paradigm: Deeply couples planning with atomic skill abstractions—the output of the "thinking" phase directly drives MoE routing decisions.
Noise-scheduled Skill Embedding: Adopts diffusion model concepts to map discrete skill labels to a continuous embedding space for effective routing.
Additive Continual Learning: Solves the forgetting problem through architectural isolation (freezing old experts) rather than complex regularization.
Physically Grounded Data Generation: Uses principal axis motion analysis to provide more reliable action segmentation than pure VLM video interpretation.

Limitations¶

Skill label quality depends on InternVideo2.5, which may struggle with rare or highly specialized operations.
Top-1 routing may be insufficient for tasks requiring simultaneous "multi-skill" coordination (e.g., pushing while rotating).
Linear growth of parameters with each new skill expert lacks a mechanism for expert merging or pruning.
Improvement on CALVIN is smaller than on LIBERO, likely due to a misalignment between CALVIN task granularity and defined atomic skills.

vs. \(\pi_0\)/\(\pi_{0.5}\): These models predict actions via pure flow matching without explicit planning. AtomicVLA adds thinking and routing, yielding a +10% Gain on LIBERO-Long.
vs. SayCan/Inner Monologue: These decouple planning (LLM) from execution (controller). AtomicVLA unifies them within a single model to close the modality gap.
vs. MoDE: While MoDE uses denoising timesteps for routing, it remains token-level. SG-MoE uses semantic skill labels, proving skill-level routing is more effective (+5.7%).

Rating¶

Novelty: ⭐⭐⭐⭐⭐
Experimental Thoroughness: ⭐⭐⭐⭐
Writing Quality: ⭐⭐⭐⭐
Value: ⭐⭐⭐⭐⭐