Recoleta Item Note

SkillVLA: Tackling Combinatorial Diversity in Dual-Arm Manipulation via Skill Reuse

vision-language-actiondual-arm-manipulationskill-reusecombinatorial-generalizationrobot-foundation-model

Summary

SkillVLA studies the problem of "combinatorial diversity" in bimanual robots: many dual-arm tasks are actually different combinations of left- and right-hand single-arm skills, but existing VLAs usually entangle the actions of both arms, making it difficult to recombine learned skills. The paper proposes a hierarchical, skill-adaptive bimanual VLA that enables the robot to reuse single-arm skills and activate bimanual collaboration only when needed.

Problem

Existing bimanual VLAs often directly predict concatenated joint actions, which causes left- and right-arm skills to become entangled and limits the model to reproducing action pairings seen during training.
Bimanual tasks exhibit clear combinatorial explosion: if the left and right hands each learn multiple skills, the number of possible combinations can grow approximately quadratically, making it impossible to teach and learn them one by one.
This matters because if a general-purpose bimanual robot cannot flexibly recombine existing single-arm skills, it will struggle to scale to new tasks, long-horizon tasks, and continual learning scenarios.

Approach

Proposes a two-level reasoning framework: a high-level VLM first uses visual/language input to generate subtask text prompts separately for the left and right arms, explicitly decomposing what each arm should do.
At the low level, separate action-generation streams are built for the left and right arms, each producing actions based on its corresponding prompt and observations, thereby supporting direct recombination of learned single-arm skills under novel left-right pairings.
For tasks that truly require tight coordination, it introduces adaptive cross-arm communication: the two action streams exchange information through cross-attention, but the communication strength is controlled by a cooperation estimator.
The cooperation estimator outputs a scalar lpha∈[0,1] (denoted as alpha in the paper), and learns when bimanual coupling is needed and when decoupling should be maintained by comparing behavior cloning errors with communication turned "on" versus "off."
It further incorporates VLM priors, temporal smoothing, and conservative communication regularization, and can discretize the cooperation gate to 0/1 to improve inference stability.

Results

On 9 unseen skill recombination tasks, SkillVLA achieves an average success rate of 0.51, while π0.5 = 0.0, π0-FAST = 0.0, and TwinVLA = 0.04, showing clear combinatorial generalization to unseen left-right skill pairings.
On specific recombination tasks, SkillVLA scores Cup×Cake 0.7, Cup×Stir 0.4, Cup×Smash 0.5, Box×Cake 0.6, Box×Stir 0.4, Box×Smash 0.5, Mug×Cake 0.6, Mug×Stir 0.3, Mug×Smash 0.6; the main baselines are nearly all 0.0.
On learned skills evaluation, SkillVLA averages 0.78, comparable to π0.5 at 0.77, and better than π0-FAST at 0.70 and TwinVLA at 0.67, indicating that its improved recombination ability does not significantly harm performance on known skills.
The paper also claims performance competitive with baselines on three categories of highly cooperative tasks, suggesting that adaptive communication can still represent tightly coordinated bimanual behavior, though the excerpt does not provide specific numbers.
In two multi-stage long-horizon tasks, SkillVLA can automatically determine cooperation requirements by stage, and reduces execution time by 21% through parallelizable recomposable skills; the excerpt does not provide the full baseline table values.
The paper further claims that in continual learning with limited demonstrations, skill reuse can significantly help acquire new skills, but the excerpt does not provide quantitative results.

Link

http://arxiv.org/abs/2603.03836v1

Built with Recoleta

Run your own research radar

Turn arXiv, Hacker News, OpenReview, Hugging Face Daily Papers, and RSS into local Markdown, Obsidian notes, Telegram digests, and a public site.

View repo 5-minute quickstart