Robot manipulation policy stops leaning on pretraining alone: Axiom-0 reaches 62.32% success on RoboCasa365
Synopsis
Axiom-0 uses hierarchical self-refinement to improve both the vision-language model's internal representations and the conditioning before action generation, reaching 62.32% overall success on RoboCasa365, 4.22 percentage points above the highest published leaderboard score of 58.1% from Paimon-0.
Interpretation
Task completion on unfamiliar task compositions rises to 62.32% while inference overhead grows only modestly. The highest previously published leaderboard score was Paimon-0 at 58.1%, with Xiaomi-Robotics-1 as a reference around 57.4%. On RoboCasa365, Axiom-0 achieves 62.32% overall success, 4.22 points above Paimon-0 and 4.92 points above Xiaomi-Robotics-1.
The policy connects pretrained multimodal understanding with learned control and recursively updates the conditioning before generating actions. Earlier vision-language-action models used pretrained representations directly, without recursively updating conditioning before action generation. Under the evaluated profiling setup, mean action-chunk generation latency increases by 1.7% over Xiaomi-Robotics-1.
Perspective
The result is evaluated on the RoboCasa365 benchmark and is aimed at researchers and engineers studying generalization of vision-language-action policies. Evaluation covers manipulation success on unfamiliar task compositions and reports the change in action-chunk generation latency.
The 1.7% inference overhead increase is measured under the evaluated profiling setup; latency changes under other hardware or batching configurations are not stated. The success-rate gain comes from evaluation on this benchmark and does not cover real-robot deployment.