cs.LGNov 12, 2026

A Recipe for Physical Action Model Training

Authors: Shaojiang Wang, Geer Yang

Organizations: Institute of AI for Industries, Chinese Academy of Sciences, Nanjing, China · Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · University of Chinese Academy of Sciences, Beijing, China

Abstract

Choosing robot model capacity and demonstration budgets also requires decisions about data composition, representation, optimization, and deployment. We connect these decisions through a common conditional-distribution target. Nested learning oracles give exact risk accounting; local information geometry, conditional mutual information, and empirical convergence provide conditional training prescriptions. Known-target numerical checks expose capacity-data coupling, representation-induced mixture mismatch, and an example where deployment KL increases while negative log-likelihood decreases. An analysis of 206 public PushT trajectories evaluates fixed-feature information-design criteria without fitting a policy. Under asymmetric evaluation targets, the selected allocations reduce this criterion by 4.5-6.2% relative to target matching; balanced targets show no benefit. These results clarify which quantities can support training recommendations and when an apparently favorable predictive metric has a different interpretation.

Figures & tables

Explore similar work

Sep 19, 2026

Robot World Models Are Not Invariant to How the Actions Are Written

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.
Oct 14, 2026

Human Demonstrations as Post-Training Data for Any Robot Body: Retargeting the Object, Not the Hand

Human demonstrations are today used largely for pretraining vision-language-action models, with little ability to use the same data for target-embodiment or task-oriented training, which scalable deployment of robots in industrial and general environments depends on. A human demonstration recorded on a simple and scalable device such as a mobile phone can serve as post-training data for a robot whose body differs in structure and capability from the human who recorded it. This is made possible not by retargeting the hand's motion but the object's. A capture yields a body-free record: the object's pose over time, its contact intervals, and a reconstructed 3D object with provenance. Any robot, given as a URDF and a short descriptor, is solved against that record and scored in simulation against the demonstrated object motion as ground truth: the human-to-simulation leg of human-to-simulation-to-robot transfer, measured per robot. Each (robot, episode) pair gets one outcome: pass, failure with a named mode, or refusal, with its reason, before any motion. The method runs as a working system over a growing collection; the set analysed here is a sample of it: 42 captures solved on 8 robots, 157 pairs, of which 31.2% pass, 52.2% fail, 14.0% are refused and 2.5% are unauditable (kinematic feasibility, +/-50 mm / +/-15 deg, in simulation). Refusals held up: every pair the checks refuse on stored evidence yet ran anyway failed (13 of 13; 5 on a body whose public model welds its hand). A description file could not be shown to screen a robot: its outcomes were barely more predictable than guessing. No policy has yet been trained on the record. The record decides what a robot may post-train on; whether the policy then works on the real robot is the simulation-to-robot question, answered downstream.
Oct 26, 2026

Behavioural Quality Is Not Learning Utility: Objective Mismatch in Human Behaviour Models for Robot Learning

Human behaviour models are increasingly used to provide simulated human interactions for training robots via reinforcement learning (RL). Existing approaches primarily evaluate these models according to how closely their generated behaviours resemble real human data, while their usefulness as training partners is less commonly assessed. This creates a potential objective mismatch: a model that performs well according to behavioural evaluation metrics may not necessarily provide useful training experiences for downstream robot policy learning. Across two human-robot interaction (HRI) domains, highway driving and object handover, we evaluate different types of human behaviour models using common generative-model evaluation metrics. We consider multiple human trajectory representations and compare the resulting evaluations with the downstream performance of RL policies trained using the corresponding models. We find that existing behavioural evaluation metrics do not consistently predict downstream policy performance across domains and trajectory representations, while their evaluations are sensitive to the choice of trajectory representation. We also show that even when the true testing behaviour distribution is known, training against a different distribution may yield better downstream performance than training directly against the true distribution.