Event
Study: pi0.5 actions bind to absolute coordinates
2 reports · 2 sources
Researchers including Bryce Grant released an interpretability study of vision-language-action (VLA) models, reporting that when activations from task A were injected into pi0.5's action expert while the policy viewed task B's scene, 99.6% of episodes (n=1,968) executed task A's motion trajectory, indicating that the injected program binds to absolute workspace coordinates and ignores the visible scene. On a physical UR10e, targeted replay caused the instructed task to be overwritten 26/26 times in one direction and 18/26 times in the other; no-injection and random-direction controls were both at the floor. After the source object was moved 10–15 cm, the robotic arm still grasped and placed according to the recorded coordinates in 50/50 trials, grasping the wrong block when an incorrect block was present and grasping at empty space when no block was present. The team also open-sourced the Action Atlas platform, covering more than 520 trained SAEs, 79 identified concepts, and UR10e scored replay; the same intervention was applied to six VLAs on LIBERO, MetaWorld, SimplerEnv, and ALOHA, totaling more than 420,000 replays.
Report timeline
- Activation injection study dissects vision-language-action models
Researchers propose an interpretability pipeline to analyze what a fine-tuned vision-language-action (VLA) policy is computing internally. On pi0.5, injecting task A's activations into the action expert while the policy views task B's scene caused 99.6% of episodes (n=1968) to execute task A's motor trajectory, indicating the injected program is bound to absolute workspace coordinates and ignores the visible scene.
arXiv cs.RO ·
- Mechanistic Study of VLA Models: Not All Features Are Created Equal
The study proposes an interpretability pipeline and finds that injecting task A's activations into the action expert causes pi0.5 to execute task A's motor trajectory in 99.6% of episodes (n=1,968), and that the injected program is bound to absolute workspace coordinates while ignoring the visible scene, offering a mechanistic explanation for the perturbation brittleness reported by benchmarks.
HuggingFace Daily Papers ·