Can Predicted Dynamics Exist in the Physical World?

Hugging Face Daily Papers Papers

Summary

This paper proposes a method for validating physical admissibility of AI predictions by using a prediction-control interface with kinematic and dynamic conditions. It demonstrates high accuracy in filtering invalid proposals on the LeRobot PushT task.

Predictive Physical AI systems output state rollouts, action chunks, and latent plans, yet a low root-mean-square error (RMSE) does not imply that a particular proposal is physically executable. We formulate physical admissibility as a prediction-control interface: before execution, a decoded proposal is treated as candidate dynamics and evaluated using kinematic, dynamic, and direct-to-composed horizon conditions. Passing is not a certificate of task success; rejection identifies violation of the specified physical envelope and gives a component-level reason. On Hugging Face LeRobot PushT, controlled falsification shows that one-step prediction-RMSE and standardized dynamics residuals reach area under the receiver operating characteristic curve (AUC) 0.982 and 0.972, kinematic-only conditions reach AUC 0.592, and the full gate reaches AUC 0.957 with condition-level attribution. In replay-based intervention experiments, residual-based filters and the full physical-admissibility gate prevent 87-$89% of invalid proposals while preserving mean progress near 0.998.
Original Article
View Cached Full Text

Cached at: 06/02/26, 03:36 PM

Paper page - Can Predicted Dynamics Exist in the Physical World?

Source: https://huggingface.co/papers/2606.00089

Abstract

Physical admissibility validation for AI systems uses prediction-control interfaces with kinematic and dynamic conditions to filter invalid proposals while maintaining high performance.

Predictive Physical AIsystems outputstate rollouts,action chunks, andlatent plans, yet a lowroot-mean-square error(RMSE) does not imply that a particular proposal is physically executable. We formulatephysical admissibilityas aprediction-control interface: before execution, a decoded proposal is treated as candidate dynamics and evaluated using kinematic, dynamic, anddirect-to-composed horizon conditions. Passing is not a certificate of task success; rejection identifies violation of the specified physical envelope and gives a component-level reason. On Hugging Face LeRobot PushT, controlled falsification shows that one-step prediction-RMSE and standardized dynamics residuals reach area under thereceiver operating characteristic curve(AUC) 0.982 and 0.972, kinematic-only conditions reachAUC0.592, and the full gate reachesAUC0.957 with condition-level attribution. In replay-based intervention experiments,residual-based filtersand the full physical-admissibility gate prevent 87-$89% of invalid proposals while preserving mean progress near 0.998.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2606\.00089

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.00089 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.00089 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.00089 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

A Path-Space Formulation of Prediction in World Models: From a Single Action to Prediction, Planning, and Irreversibility

arXiv cs.LG

This paper proposes a path-space formulation of prediction in AI world models, treating the distribution over future trajectories as the fundamental predictive object. It shows that prediction, planning, and uncertainty emerge as operations on a single action functional, and demonstrates that attention asymmetry in learned models correlates with irreversibility in the data.

Sim-to-real transfer of robotic control with dynamics randomization

OpenAI Blog

OpenAI researchers demonstrate a method to bridge the reality gap in robotic control by training policies with randomized simulator dynamics, enabling robots trained purely in simulation to successfully transfer to real-world tasks like object manipulation without physical training.

Anticipatory Reinforcement Learning for Trajectory Tracking

arXiv cs.LG

This paper introduces a predictive formulation for deep reinforcement learning that augments the state space with future reference horizons to enable anticipatory control for trajectory tracking. Simulation results show significant error reduction, though zero-shot transfer to physical hardware reveals a sim-to-real gap.