[Project SkillVLA] Breaking the Bimanual Bottleneck: Solving Combinatorial Diversity through Skill Reuse

TL;DR

Dual-arm robots are often limited by the "combinatorial explosion" of tasks; learning every possible combination of left and right hand actions is impossible. SkillVLA introduces a hierarchical reasoning framework that allows robots to learn single-arm skills and then zero-shot recombine them for new tasks. By shifting from monolithic joint-action prediction to a gated, disentangled architecture, it improves success rates in unseen bimanual combinations from 0% (SOTA baselines) to 51%.

The Core Challenge: Skill Entanglement

Current Vision-Language-Action (VLA) models treat bimanual control as a single, high-dimensional action vector. While this works for tasks seen in training, it leads to Skill Entanglement.

If a robot only learns to "pick a cup" with the left hand while the right hand is idle, it "entangles" the cup-picking behavior with right-hand idleness. When asked to "pick a cup" while the right hand "stirs a pot," the model fails because it has never seen that specific joint distribution. This is a geometric failure: the target action lies entirely outside the training manifold.

Methodology: Disentangled Reasoning & Adaptive Cooperation

SkillVLA tackles this by decoupling both the reasoning and the execution.

1. Two-Level Reasoning

Instead of one prompt for two arms, a High-Level VLM generates two distinct sub-prompts (e.g., $u_L$: "Open drawer", $u_R$: "Pick apple"). These are fed into independent Low-Level VLMs and Action Experts. This ensures that the latent representations for each arm are not "polluted" by the other's state during independent phases.

2. The Collaboration Estimator (The Gating Magic)

Not all tasks can be independent. Lifting a heavy box or passing an object requires tight coordination. SkillVLA includes an Adaptive Cross-Attention mechanism controlled by a scalar $\alpha \in [0,1]$.

• $\alpha \approx 0$: The arms operate as independent agents (Parallelism).
• $\alpha \approx 1$: Information flows between action experts (Coordination).

Figure 1: The SkillVLA framework showing the two-level reasoning and the gated collaboration mechanism.

Experimental Breakthroughs

The authors validated SkillVLA on a real dual-arm setup across single-arm reuse, cooperative tasks, and long-horizon scenarios.

Zero-Shot Recomposition

When tested on 9 unseen pairings of learned skills, monolithic models like π0.5 and π0-FAST achieved a 0% success rate—they simply froze or reverted to training patterns. SkillVLA averaged 51%, proving that its architecture successfully recovers the product space of independent skills.

Efficiency in Long-Horizon Tasks

In tasks like "Tubes" (moving tubes then lifting a rack), SkillVLA identified that the first stage can be done in parallel. While baselines moved one arm at a time (sequential execution), SkillVLA used both arms simultaneously, reducing execution time by 21%.

Table 1: SkillVLA's decisive advantage in zero-shot recomposition tasks compared to baselines.

Continual Learning: Faster Adaptation

SkillVLA also shines in acquiring new dual-arm skills. Because it reuses existing single-arm "knowledge," it can learn a cooperative task (like "Shake Cup") with significantly fewer demonstrations. It effectively uses the single-arm skill as a "strong prior," only needing to learn the offset or the coordination via the cross-attention module.

Critical Insight & Future Outlook

The success of SkillVLA suggests that compositionality is the key to scaling robot learning. Rather than building bigger monolithic models, we should build models that can factorize the world into reusable primitives.

Limitations: The model currently relies on natural language as the "bridge" between high and low levels. While intuitive, language can be ambiguous. Future iterations might benefit from more abstract "skill embeddings" or latent goal vectors to handle even more granular coordination.

Conclusion

SkillVLA marks a shift from "black-box" bimanual control to a "structured-reasoning" approach. By explicitly modeling when to cooperate and when to act independently, it allows dual-arm robots to finally escape the constraints of their training data and embrace the combinatorial diversity of the real world.