Source-linked AI summary
AtomicVLA: Unlocking the Potential of Atomic Skill Learning in Robots
Likui Zhang, Tao Tang, Zhihao Zhan, Xiuwei Chen, Zisheng Chen, Jianhua Han, Jiangtong Zhu, Pei Xu, Hang Xu, Hefeng Wu, Liang Lin, Xiaodan Liang
TL;DR
Existing VLAs struggle with long-horizon reasoning, continual skill acquisition, and interference among learned skills. AtomicVLA unifies planning and execution through atomic skill abstractions and SG-MoE experts, achieving improvements across simulation and real-world evaluations.
Problem
Existing VLAs face challenges in real-world long-horizon tasks and continual skill acquisition, while incremental learning can cause interference and catastrophic forgetting.
Method
AtomicVLA jointly generates task plans, atomic skill abstractions, and fine-grained actions using a unified framework with an expandable SG-MoE skill library.
Results
AtomicVLA outperforms baselines across simulation and real-world long-horizon and continual-learning evaluations, including a 2.4% average improvement on LIBERO.
Takeaways & Limitations
Atomic skill abstraction and dynamic expert composition support long-horizon execution and lifelong robotic skill expansion within the evaluated settings.
Takeaways & Limitations
CALVIN’s evaluation rules exclude successful recoveries after failures, so reported metrics may slightly underestimate the model’s capability.
Abstract
from arXiv · showhide
Recent advances in Visual-Language-Action (VLA) models have shown promising potential for robotic manipulation tasks. However, real-world robotic tasks often involve long-horizon, multi-step problem-solving and require generalization for continual skill acquisition, extending beyond single actions or skills. These challenges present significant barriers for existing VLA models, which use monolithic action decoders trained on aggregated data, resulting in poor scalability. To address these challenges, we propose AtomicVLA, a unified planning-and-execution framework that jointly generates task-level plans, atomic skill abstractions, and fine-grained actions. AtomicVLA constructs a scalable atomic skill library through a Skill-Guided Mixture-of-Experts (SG-MoE), where each expert specializes in mastering generic yet precise atomic skills. Furthermore, we introduce a flexible routing encoder that automatically assigns dedicated atomic experts to new skills, enabling continual learning. We validate our approach through extensive experiments. In simulation, AtomicVLA outperforms $π_{0}$ by 2.4\% on LIBERO, 10\% on LIBERO-LONG, and outperforms $π_{0}$ and $π_{0.5}$ by 0.22 and 0.25 in average task length on CALVIN. Additionally, our AtomicVLA consistently surpasses baselines by 18.3\% and 21\% in real-world long-horizon tasks and continual learning. These results highlight the effectiveness of atomic skill abstraction and dynamic expert composition for long-horizon and lifelong robotic tasks. The project page is \href{https://zhanglk9.github.io/atomicvla-web/}{here}.
1. Introduction
AtomicVLA addresses the difficulty of deploying VLAs on long-horizon tasks and continually acquiring skills by unifying planning with fine-grained execution and organizing skills into expandable experts.
- Current VLAs struggle with complex long-horizon tasks and continual acquisition of new skills in real-world deployments.
- Two-stage planner-controller designs can lack mutual awareness, while incremental fine-tuning demands resources and can cause catastrophic forgetting.
- AtomicVLA unifies task planning and action execution by adaptively generating natural-language instructions or latent actions.
- SG-MoE builds a scalable atomic-skill library with shared and dedicated experts, while an extensible router supports continual skill expansion.
- 2.4% average improvement was achieved on LIBERO, 10% on LIBERO-LONG, and real-world experiments reported 18.3% and 21% improvements for long-horizon and continual learning tasks.
2. Related Work
Related work establishes VLAs as general-purpose robotic policies but identifies limitations in structured planning, modular coordination, and continual learning. AtomicVLA reframes MoE around reusable atomic skills with dynamically expandable experts.
- VLAs fine-tune pretrained VLMs on robotic datasets to map visual and linguistic inputs directly to motor actions.
- External planners can decompose long-horizon tasks, but separate controllers may produce misaligned decisions because action, vision, and language are not unified in one latent space.
- MoE improves specialization through conditional computation, while prior robotics applications often treat experts as interchangeable architectural components.
- AtomicVLA constructs an expandable library in which experts correspond to semantically meaningful, reusable atomic action primitives.
- Continual-learning research has explored skill discovery, but dedicated continual-learning investigations for VLAs remain limited.
3. Method
AtomicVLA integrates adaptive thinking and acting with skill-guided expert routing. Its atomic-skill embeddings, sparse expert composition, continual expansion mechanism, and trajectory-based decomposition support precise execution and scalable learning.
- Inference pipeline: During inference, the policy predicts a modality, updates the atomic skill during thinking, routes the skill during acting, and executes the resulting action.
- Adaptive thinking and acting: The model alternates between thinking, which produces a task plan and atomic skill abstraction, and acting, which generates executable actions.
- Atomic skill embedding: Each atomic skill is mapped through a scalar noise level and embedding function into a high-dimensional vector that supports semantic separation and expert routing.
- Skill-guided mixture of experts: SG-MoE combines a skill router, a shared π0 expert, and dedicated atomic-skill experts, activating the top-scoring skill expert for execution.
- Skill-guided mixture of experts: The final action chunk combines shared and selected-expert outputs, retaining π0 generalization while improving fidelity for specific skills.
- Continual learning with skill expansion: New skills expand the library by adding an expert and extending the router, with copied router weights and small random branch initialization enabling minimal fine-tuning.
- Embodied data generation: Trajectory decomposition uses translational, rotational, and gripper-state changes to produce temporally precise and semantically interpretable atomic-action boundaries.
4. Experiments
AtomicVLA is evaluated on LIBERO, CALVIN, and real-world Franka tasks spanning simulation, long-horizon manipulation, mixed-task execution, and continual skill expansion. Across these settings, it improves benchmark performance, supports error recovery, and reduces interference during new-skill learning.
- Experiments Setup: AtomicVLA is evaluated on LIBERO and CALVIN, including all four LIBERO suites and the CALVIN ABC-D split for long-horizon planning.The training setup uses pretrained π0 and π0.5 foundations, with five skill experts for LIBERO and real-world experiments and eight for CALVIN.
- Results on Simulation: 96.6% average success on LIBERO exceeds the strong baseline by 2.4%, while LIBERO-LONG reaches 95.2%, a 10% improvement over π0.AtomicVLA* reports 97.8% average success and 96.2% on LIBERO-LONG.
- Results on Simulation: AtomicVLA detects execution anomalies, regenerates atomic skill abstractions, and resumes task execution after failures.The reported CALVIN metrics may slightly underestimate this capability because recoveries after failures do not count as valid completions.
- Results on Simulation: 4.09 average task length on CALVIN exceeds π0 by 0.22, while AtomicVLA* reaches 4.27, exceeding π0.5 by 0.25.AtomicVLA* also improves completion rates by 5.8%, 6.2%, and 6.6% on the final three evaluation stages.
- Results on Real-world Robot: 20% and 18.3% improvements over the baseline are reported for AtomicVLA and AtomicVLA*, respectively, in mixed-training real-world long-horizon tasks.AtomicVLA* reliably completes configurations that π0.5 fails, with a clearer advantage on door-closing tasks.
- Results on Real-world Robot: AtomicVLA* maintains stable performance after continual learning and achieves a 21% overall improvement across five tasks compared with π0.5.The baseline π0.5 loses approximately 15% on average, including a 20% decrease on the stack task.
5. Conclusion
AtomicVLA unifies task planning and action execution for long-horizon tasks and continual skill expansion. Its scalable atomic-skill architecture supports new capabilities while mitigating skill interference and forgetting.
- AtomicVLA unifies task planning and action execution for long-horizon tasks and continual skill expansion.
- The framework adaptively decides task plans and generates latent action outputs through a unified architecture.
- SG-MoE and a specialized skill router construct an atomic skill-guided expert library.
- New skills can be acquired by extending the skill router and adding corresponding skill experts.
- Experiments in simulated and real-world environments demonstrate strong performance in long-horizon tasks and continual learning.
- AtomicVLA mitigates skill interference during joint training and alleviates forgetting and performance degradation during continual acquisition.
A.1. Video Demonstration
The appendix points readers to an attached video file for a quick overview of AtomicVLA. The passage provides no further description of the demonstration.
- The appendix directs readers to an attached video file for a quick overview of AtomicVLA.
- The referenced material is presented as a quick overview rather than a detailed experimental description.
- No experimental findings are stated in this passage.
A.2. Future Work and Limitations
The paper identifies unresolved limitations in skill decomposition, planning fidelity, and the broader exploration of multi-skill VLA learning. It also points to reinforcement learning and richer data as possible directions for extending AtomicVLA.
- AtomicVLA relies on accurate atomic skill abstractions and well-trained skill experts, leaving several advantages insufficiently explored.
- The skill router depends on VLM reasoning and planning fidelity to produce accurate atomic skill abstractions during execution.
- Large-scale web data combined with embodied experience may improve VLM skill decomposition, task planning, and expert-library quality.
- The work investigates skill interference from multi-skill joint training through controlled experiments.
- Reinforcement learning with a pretrained skill-expert library may help AtomicVLA generalize to novel tasks under few-shot or zero-shot settings.
A.3. Additional Details
Additional details describe AtomicVLA’s data preparation, evaluation settings, error-recovery behavior, and performance across simulation and real-world experiments. Results show strong broad performance alongside distribution-shift and task-specific weaknesses.
- Additional Details: LIBERO data are organized into five atomic actions: Pick, Place, Open, Close, and Turn, with undersampled categories upsampled for balance.
- Additional Details: CALVIN data use eight atomic abstractions and an eight-expert skill library, with trajectories capped at 64 frames.
- Additional Details: Atomic action annotations use principal-axis analysis with thresholds of 3 cm, 0.05 radians, and 0.1 gripper change.
- Additional Results: On CALVIN, AtomicVLA* achieves success rates close to 100% on most tasks but only 20–30% on several Push blocks right tasks.
- Additional Results: 43.3% average accuracy on complex scenes is 10% higher than π0.5, while corn picking improves by 15%.With 12 experts, inference latency is 160 ms on one H20 GPU.
- Additional Results: AtomicVLA demonstrates simulated robustness, real-world error recovery, and generalization across simple-to-complex tasks and regular-to-irregular objects.