Skip to main content
Back to timeline
arXivSource publication:

Imitation policies learn a proprioceptive anchor: one measurement predicts sensitivity at new noise levels, but the protection fails in deployment

Synopsis

Imitation policies trained on noisy object-position readings do not discount the reading by the classical reliability ratio. They combine it with a proprioceptive anchor, weighting the reading by the ratio of reading noise to the anchor's error variance, and that variance belongs to the learner and is far smaller than regression on the demonstrations implies. Measured once, the law predicted ten held-out sensitivities across two robomimic tasks and two learner classes with mean absolute error 0.025, against 0.37 for classical errors-in-variables regression. In closed loop the policy produces its own state, and none of the 15 policies that passed the registered gates kept its open-loop protection against a persistent reading offset, though 8 followed the offset fully.

Source-provided article image: Imitation Policies Learn a Proprioceptive Anchor: A Predictive Law for Perception Noise and Its Closed Loop Cost
arXiv

Interpretation

For an imitation policy trained on noisy object-position readings, how much the policy relies on the reading can be predicted by a quantitative law whose key parameter comes from the policy itself rather than from the demonstrations. Classical errors-in-variables regression writes the discount as the ratio of true-input variance to noise variance, and from that predicts that at 12 mm of noise a policy should keep 86% of its clean sensitivity. On the Square task with diffusion policies, policies trained at 12 mm kept only 23% of their clean sensitivity on the informative axis; the classical ranking held but the magnitude was far off.

One measurement extrapolates: a weight measured on a 12 mm calibration policy predicts the same learner's response to the reading at other noise levels. No quantitative account previously predicted this discount for modern policy classes before training, and classical errors-in-variables regression was the only candidate and failed. Across two tasks, two learner classes and four new noise levels, the anchor law's ten held-out predictions had mean absolute error 0.025 with all ten inside their registered tolerances, against 0.37 and none inside tolerance for classical errors-in-variables regression.

The discount comes from a proprioceptive anchor: demonstrators steer toward the object, so the robot's own state carries information about where the object is, and the learner combines anchor and reading by reliability. Classical errors-in-variables regression estimates the anchor's error from the demonstrations, and every linear and nearest-neighbour model puts it near 30 mm. Trained policies act as if the anchor error were about 7 mm for diffusion and about 15 mm for the MLP, and the bias at clean readings is small, ruling out the alternative that policies merely act as if the noise were larger.

Open-loop protection does not survive closed loop, because the policy produces its own state and a persistent reading offset becomes indistinguishable from a displaced object. The registered alternative was that the anchor keeps protecting in closed loop, holding sensitivity to a persistent offset near zero. None of the 15 policies that passed the registered gates kept its open-loop protection and 8 followed the offset fully; for the MLP the offset response matched the response to a real object displacement with r = 0.94 (post hoc).

Perspective

The results come from the robosuite simulator with synthetic Gaussian noise added to the object position, low-dimensional observations, 200 human demonstrations per task and two seeds per configuration; sensitivity is probed only on the translational action at five early time steps before any grasp. Whether the law holds for image-based policies and real perception noise is untested, and the anchor's error variance must be measured from the policy rather than derived from the data alone.

All noise is Gaussian and, except in the benchmark's AR(1) conditions, independent across time steps; real perception errors depend on viewpoint, occlusion and object pose and can be heavy-tailed. The anchor's error variance must be measured rather than derived, and it depends on the amount of data: the 12 mm MLP's ratio was 0.362, 0.500 and 0.602 at 50, 100 and 200 demonstrations. On Square's x axis, where the object position varies by only 1.4 mm, diffusion policies still keep about 0.37 sensitivity, where a calibrated anchor would allow at most 0.05. In the prospective success test, open-loop sensitivity and training noise are confounded, and a curve interpolated in training noise predicts success as well. The closed-loop offset test's three registered predictions were mixed for both learners, and the real-displacement check that explains the in-between policies was run for the MLP only.