Robot world models fail when action commands are rewritten differently
Robot World Models Are Not Invariant to How the Actions Are Written
RoboticsMachine Learning
Summary
Robots learn to perform tasks by following action commands, which can be expressed in different ways. This paper shows that changing how these commands are written—either as exact target positions or as changes from the current position—can confuse the robot's internal model, leading to much worse task performance. The authors demonstrate that the same planned movements look very different to the robot's predictive model when expressed in different formats, even though all the necessary information is still there. They propose a method to fix this by averaging across the two command forms during training, improving results but not fully solving all disagreements.
What this means in practice
- •For robotics software engineers: Develop robot control systems that remain reliable regardless of how action commands are formatted by integrating model training across multiple action parameterizations.
- •For industrial automation teams: Improve robustness of robot task planning in factories by mitigating failures caused by differing robot command styles in predictive models.
Authors
Ahmed Karim, Leon Chlon
Abstract
A robot policy is trained with one of two action parameterizations: absolute joint targets, or deltas relative to the current state. The choice is a live engineering decision in robot learning, and a world model conditioned on actions inherits it silently. We show the inheritance is catastrophic. A latent dynamics model trained on one parameterization and handed the identical commanded trajectory written in the other collapses: retrieval degrades by 2.6-13.4x across three robot datasets and two morphologies, goal-conditioned action selection falls from 53% to 15%, and on PushT the two beliefs about the same future are near-orthogonal (cos = 0.067, worst case -0.377), so the predictor does not degrade gracefully, it answers a different question. This is not a distribution-shift artifact in the usual sense: the two encodings are mutually reconstructible at R^2 = 0.996 given the joint input, so no information is lost, and we give the test that separates a valid re-parameterization from a lossy summary or a sensor swap. The test rejected three of the four axes we proposed. The defect lives in the action channel, which the invariance literature for visual models does not examine: work there concerns crops, jitter and camera pose, while the parameterization of the commands goes unaudited. The repair is averaging over the two encodings, and where it goes matters. Averaging the objective restores task performance by itself; averaging the outputs, safe for probabilities by concavity, is not available for direction-valued prediction, where the normalized mean can score below every member of the orbit. What objective-averaging leaves behind is the tail: worst-case agreement stays at 0.78, a disagreement penalty closes it to 0.995, and over a latent rollout it is the difference between a worst case that erodes and one that holds. On PushT, averaging alone does not repair the axis.