Summary
StellaVLA conditions a VLA on a single retrieved demonstration at test time rather than fine-tuning on it. An automated offline pipeline converts raw trajectories — from either robot teleoperation or human/XR demonstrations — into “structured demonstrations” (task plan, sub-goal descriptions, verbalized 3D motion) at zero human-annotation cost, which the policy then conditions on in-context.
Key Contributions
- A structured-demonstration representation (plan + subgoals + verbalized 3D motion) built automatically from raw trajectories, avoiding manual annotation while giving the policy richer conditioning than a raw video clip.
- In-context conditioning that lets a single retrieved demonstration steer generalization, without gradient updates or fine-tuning at deployment time.
- Reports leading the VLA-Arena leaderboard (overall score 0.63 vs. 0.44 for π0.5 and 0.22 for LingBot-VLA as of 2026-08-01) and strong results on LIBERO (98.8%) and LIBERO-Plus (85.1%).
Strengths
- Zero-annotation-cost structuring is a meaningful practical advantage over methods that require hand-labeled subgoals or language annotations for in-context conditioning.
- Supporting both robot and human/XR demonstration sources widens the pool of usable conditioning examples considerably.
- Strong, directly comparable benchmark numbers against named competing baselines (π0.5, LingBot-VLA) give a concrete sense of the improvement, not just an isolated claim.
Weaknesses
- Performance is contingent on retrieval quality: if the retrieved demonstration is a poor match for the deployment scenario, the paper’s coverage does not detail how gracefully the method degrades.
- LIBERO and LIBERO-Plus are simulation benchmarks; the extent of real-robot validation is not clear from available coverage, which matters for a method whose main claim is deployment-time generalization.
Open Questions
- How large and diverse does the demonstration retrieval pool need to be before the in-context approach saturates in benefit?
- Does structured-demonstration conditioning hold up on long-horizon, multi-stage tasks where subgoal decomposition itself becomes harder to get right automatically?
Significance
Part of a fast-growing cluster of 2026 work (alongside WIZARD, Retrieval-VLA, and Retrieve-then-Steer already in this vault) treating in-context conditioning as a lighter-weight alternative to fine-tuning for VLA generalization — notable here for its automated, annotation-free demonstration structuring and strong reported leaderboard position.