A team led by Guangyuan Wang describes an inference-time method for adapting pretrained robot policies to obstacles without collecting new demonstrations or retraining. It evaluates the robot’s projected path and spreads obstacle corrections across the action sequence, rather than correcting each step separately.
In tests across seven obstacle-constrained scenarios and four manipulation tasks, the researchers report 105 successful collision-free tasks out of 140, compared with 8 out of 140 using absolute-action pointwise guidance. These are reported experimental results; the abstract does not specify how the method performs outside those scenarios.
Why it matters
For robotics researchers trying to adapt pretrained policies to new obstacles, the result suggests that coordinating corrections across an action sequence may help without retraining, though performance beyond the tested scenarios is unknown.
