Imitation Policies Learn a Proprioceptive Anchor: A Predictive Law for Perception Noise and Its Closed Loop Cost
Abstract
Imitation policies for robot manipulation are often trained on noisy object positions, and they learn to discount the noisy reading. We show how much they discount it, and why. Because demonstrators steer toward the object, the robot's own state carries information about where the object is. A trained policy combines this proprioceptive anchor with the reading and weights the reading by , where is the reading noise and the anchor's error variance. belongs to the learner and is far smaller than a linear model of the demonstrations implies, which is why classical errors-in-variables (EIV) regression overestimates . Measured once, predicted ten held-out sensitivities across two robomimic tasks, two learners and four new noise levels, with mean absolute error 0.025 against 0.37 for classical EIV, and a parameter-free bias test rules out the simpler explanation that the policies just overestimate the noise. The discount does not carry over to deployment, where the policy produces its own state: closed-loop sensitivity to a persistent reading offset exceeded in all 15 policies that passed the registered gates, although the stricter registered prediction, that all would follow the offset fully, held for 8 of them. Post hoc, lower went with lower success, steeply for diffusion, and a registered prospective test confirmed this for diffusion on Can at two new noise levels: the measured predicted success before any rollout, within 0.03 (95% intervals ), although training noise predicts it equally well. Every test was registered before its data, including the classical benchmark that failed, and all other analyses are labelled post hoc; all results are in simulation with synthetic Gaussian noise.
Figures & tables
| Task | Learner | Noise (mm) | Measured | Anchor law | Classical EIV |
|---|---|---|---|---|---|
| Square | MLP | 24 | 0.315 | 0.295 | 0.597 |
| 6 | 0.841 | 0.849 | 0.959 | ||
| Square | diffusion | 24 | 0.093 | 0.071 | 0.598 |
| 6 | 0.431 | 0.525 | 0.960 | ||
| Can | MLP | 24 | 0.389 | 0.349 | 0.551 |
| 6 | 0.814 | 0.849 | 0.941 |
| Condition | History (frames) | Classical EIV | Measured |
|---|---|---|---|
| iid, 1.13 mm | 2 | 0.874 | 0.941 |
| iid, 11.99 mm | 1 | 0.584 | 0.187 |
| iid, 11.99 mm | 2 | 0.624 | 0.309 |
| iid, 11.99 mm | 4 | 0.652 | 0.490 |
| AR(1), 11.99 mm | 1 | 0.584 | 0.273 |
| AR(1), 11.99 mm | 2 | 0.582 | 0.303 |
| Explanation | What it predicts | What we found |
|---|---|---|
| A mis-estimated spread | A better model of the demonstrations fixes the prediction | Every linear and nearest-neighbour model misses alike (mean absolute error 0.23 to 0.25) |
| Local memorization | A stronger discount at training states than at fresh placements | The same discount at both (difference ); the nut’s orientation, unique to each demonstration, is not used as a lookup |
| Noise structure within a demonstration (MLP) | Noise held constant per demonstration is discounted much less | It is not: 0.660, against 0.602 for independent frames and 0.633 for 1-s correlated noise (the rule needed ) |
| Recall of training placements | No bias at training placements, a bias at fresh ones | Near-zero bias at both: at demonstration states, and along trajectories from training and fresh placements, against for plain shrinkage |
| Task | Learner | Measured bias | Anchor prediction | Noise inflation |
|---|---|---|---|---|
| Square | diffusion | ; (post hoc) | ||
| Square | MLP | (post hoc) | ||
| Can | MLP | ; within 0.008 per axis b (post hoc) | ||
| Can | diffusion | , | , |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Test | Registered (UTC) | Question | Registered verdict | Key numbers |
|---|---|---|---|---|
| PR4, arm | 2026-09-24 18:53 | Classical gain on a second toy task | Gain test PASS; overall FAIL | gain within 0.13 at 4 of 4 levels; chunking claim refuted; success ordering 0.74 (needed 0.8) |
| PR5, peg and arm | 2026-09-24 19:14 | Success predicted from the classical gain | FAIL | MAE 0.29 over 20 held-out configurations |
| PR6, peg and arm | 2026-09-24 22:16 | Same, frozen learner protocol, 14 new configurations | PASS | MAE 0.11 against 0.30 naive |
| PR7 E1 | 2026-10-01 01:30 | Does PR6 hold with four seeds? | PR6 STANDS | MAE 0.102 (95% CI 0.065 to 0.161), naive 0.312, p = 0.0008 |
| PR7 E2 | 2026-10-01 01:30 | History length and noise structure | PASS | MAE 0.079, Spearman 0.93 |
| PR7 E3 | 2026-10-01 01:30 | Robustness to the binning choice | PASS | MAE change +0.023 and +0.005 (limit 0.05) |
Robot Policy Learning
Robot policy learning develops controllers that map sensory observations, task goals, and interaction history to actions for embodied robots.
See all 398 papers