Robot models improve when actions explain intended effects
En termes simples
Arc is a recipe for helping pretrained robot models decide what to do. It adds short explanations that connect the current scene to the change the next action should cause, including what to avoid and how to tell whether progress was preserved.
The explanations are made automatically from existing robot videos and actions, then used to train the models further. In tests without training on those tasks, the adapted models succeeded more often than their original versions on both ordinary manipulation and reasoning-focused tasks.
La recherche
Existing robot foundation models already act from demonstrations, but those demonstrations usually show actions without the reasoning behind them.
It remains unclear whether a model can gain better control from the same data when it is taught to connect the current scene, the intended effect, and the action that should produce it.
Arc turns existing robot demonstrations into action-grounded causal traces, then adapts pretrained models to use them for control without task-specific training.
Arc automatically adds causal traces to existing demonstrations, recording the relevant state, intended effect, action, what to avoid, and completion. Pretrained models then use these traces while keeping their original prediction objectives.
In evaluations without task-specific training, adapted Pi_0.5 and Cosmos3-Nano-Policy beat their corresponding unmodified models on general manipulation and reasoning-focused tasks.
That cross-model result supports the recipe’s value, while the explicit-subtask control suggests coarse decomposition alone is insufficient.
The evidence is limited to the reported simulation tasks and 28 paired physical tasks, and the explicit-subtask control changes both the external output and whether the policy was adapted.
The authors’ explanation that one state-to-effect representation supports sequencing and recovery is an interpretation, not causal proof.
Analyse
Détails
- Licence
- CC BY 4.0