Physics Filtering Favors the Generalization of Robot Learning

Jindou Jia; Shixuan Han; Meng Wang; Gen Li; Zihan Yang; Sicheng Zhou; Kexin Guo; Jianfei Yang

cs.ROAug 24, 2026

Physics Filtering Favors the Generalization of Robot Learning

Authors: Jindou Jia, Shixuan Han, Meng Wang, Gen Li, Zihan Yang, Sicheng Zhou, Kexin Guo, Jianfei Yang,

Organizations: Nanyang Technological University · Beihang University · National University of Singapore · Beijing Aerospace Control Instrument Research Institute

Abstract

Living organisms exhibit extraordinary adaptability to unseen environments through their intrinsic physical structures and lifelong feedback-driven learning. Endowing robots with comparable generalization is critical for reliable operation in the real world. While recent approaches attempt to improve generalization by scaling training data, such strategies remain impractical for robotics, where collecting real-world demonstrations at the scale of large language models is prohibitively costly and slow. Contrary to this reliance on massive datasets, we show that robots can generalize effectively under dynamics uncertainties even with limited training data by leveraging a feedback mechanism, namely PhyFilter, that corrects learning outputs with physics-filtered learning residuals. PhyFilter operates as a lightweight, model-agnostic module whose parameters can be automatically optimized through an auto-learning algorithm, eliminating manual tuning and enabling seamless integration with diverse robot policies. We validate PhyFilter across four representative robotic systems, demonstrating that it enables quadruped robots to generalize to unseen terrains, payload variations, and speed ranges; drones to flight under unseen wind disturbances; aerial manipulators to achieve centimeter-level in-air capture despite wind and mass uncertainties; and acceleration differentiators to remain robust with distribution shift. These results show that physics-filtered feedback can serve as a powerful alternative to massive data scaling.

Figures & tables

Figure 1: Overview of our PhyFilter. a , The learning output of RL or SL is enhanced by a physics-informed filter, which helps narrow the large generalization gap. b , The implementation details of PhyFilter, which is analytic and convergent. Additional theoretical details are provided in the “Filtering learning residual” section in Methods. c , PhyFilter is validated across four representative examples. i , Policy learning for quadruped locomotion. A quadruped robot trained in a simplified simulator generalizes to unseen speeds, payload variations, and real-world terrains when equipped with PhyFilter. ii , Dynamical learning for drone maneuvering flight. A learning-based model captures mass variations, while PhyFilter helps handle unseen wind disturbance. iii , Dynamical learning for aerial manipulation in a pick-and-place task. A dedicated model learns drone-manipulator coupling uncertainties, with PhyFilter aiding in the handling of unseen wind disturbances and mass uncertainties. iv , Kinematics learning for acceleration perception. A neural network differentiator maps velocity sequences to real-time acceleration, and PhyFilter improves its robustness to distribution shift. We also demonstrate the performance of PhyFilter on a humanoid robot initially, which is detailed in Supplementary Fig. S6.

Figure 2: Simulation results of the quadruped example. a , Parallel training setup using only flat terrain. b , Maximum training rewards over five random seeds. The number of iteration and training environment are set as $15000$ and $256$ , respectively. c , Locomotion performance of the baseline and PhyFilter in the training environment. d , Testing performance at an unseen high speed of $2.8\ \text{m/s}$ . e , Testing performance under unseen payloads ranging from $15\%\sim 60\%$ of the robot’s $11.86\ \text{kg}$ weight. f , Testing performance across five unseen scenarios. g , Stairs, rough, wave, and gravel roads, unseen in training, are considered during deployment.

Figure 3: Experimental results of the quadruped example. Several unseen scenarios including flagstone, lawn, sand, and gravel terrains are evaluated. Notably, the policy is trained solely on simulated flat terrain with no post-deployment tuning.

Figure 4: Experimental results of the drone maneuvering example. a , The employed SEER-I is trained under mass uncertainty, while the testing scenario further includes the wind disturbance, as depicted in b . c , In tests, the drone is commanded to follow a circular trajectory with $2\ \text{m/s}$ . d , The tracking performance with only mass uncertainty under the baseline and SEER-I. e , The tracking performance with both mass uncertainty and wind disturbance under the baseline, SEER-I, and PhyFilter. f , Mean absolute tracking errors across all implemented tests. The error bar indicates the tracking errors presented in ( d ) and ( e ). g , Several flight snapshots.

Figure 5: Experimental results of the aerial manipulation example. a - c , Pick-and-place executions under the baseline, SEER-I, and PhyFilter, respectively. d , Illustration of the pick-and-place task involving coupled dynamics, wind disturbance, and mass uncertainty. e , Three-dimensional tracking performance of the compared methods. f , Tracking errors for each method.

Figure 6: Experimental results of the accelerator example. a , Training dataset containing velocities from $0$ to $2\ \text{m/s}$ . b , Testing dataset containing velocities from $0$ to $3\ \text{m/s}$ . c , Absolute prediction errors of each method on the training set. d , Absolute prediction errors on the testing set. e , Ablation study on filtering orders and parameters determined via pole-placement . Blue stars denote the auto-learned parameters, which are roughly consistent with manually optimal results. f , Distribution of the training and testing datasets. g , Ablation study across different sampling intervals. RMSE denotes the root mean square error.

Figure S1: Estimation performance of the drone example. a , Uncertainty estimation with feedback-based disturbance observer. b , Uncertainty estimation with SEER-I. c , Uncertainty estimation with the proposed PhyFilter. For reference, the ground truth of mass uncertainty is provided in ( a )-( c ), which can be deduced from the ground-truth mass in advance. It can be seen that SEER-I can only reflect the mass uncertainty, while PhyFilter can further capture other unknown wind and inherent uncertainties, leading to more accurate trajectory tracking, as shown in Fig. 4e.

Figure S2: The convergence performance of Algorithm 2 at different filter orders in the acceleration example. Each case is repeated $5$ times with different initial filter parameter $\bm{a}_{f0}$ . A higher filter order requires a smaller learning rate.

Figure S3: The trained locomotion performance of all compared methods. PhyFilter and RL2AC can yield more coordinated locomotions, whereas the baseline results in inferior performance.

Figure S4: The employed hardwares. a , The quadruped robot used in the locomotion example. b , The aerial manipulator used in the manipulation example. c , The flight drone used in the flight example.

Figure S5: Comparison between the baseline and PhyFilter under a strongly randomized training setting. a , Training scenarios at curriculum difficulty 0–0.90 (Scenes #1–#4: stairs, rough road, waves, gravel road). b , Testing scenarios at curriculum difficulty 0.92–1.00, the same terrains but beyond the training distribution. c , Testing success rates (%) across the four scenes. Even trained with full standard randomization, the baseline still fails under sufficiently out-of-distribution conditions, whereas PhyFilter consistently achieves higher success rates.

Figure S6: Humanoid robot tests under in-distribution and out-of-distribution external forces. A lateral force is applied to one arm of the humanoid, and each row shows snapshots of PhyFilter (left) and the baseline (right) at t=1, 5, and 9s. a , In-distribution test under a 20 N force: both PhyFilter and the baseline keep the robot stably standing. b , Out-of-distribution test under a 90 N force: PhyFilter maintains stable standing throughout, whereas the baseline loses balance and falls.

Training parameter	Value	Domain randomization	Range
Number of environments	$4096$	Mass	$-1\sim 3$ kg
Number of iterations	$15000$	Velocity	$-1\sim 1$ m/s
Filter order $n$	$1$	Gain $\bm{K}_{p}$	$0.8\sim 1.2$
Filter parameter $a_{0}$	$1$	Gain $\bm{K}_{d}$	$0.8\sim 1.2$
		Friction	$0.1\sim 1.25$

Table S1: Training parameters and domain randomization ranges for the quadruped example.

Robot Learning

Robot learning develops methods for robots to acquire skills and adapt from demonstrations, interaction, data, and physical models.

See all 321 papers

Oct 21, 2026

Robot Generalization Beyond Volume: The Diversity Path via Zero-Shot Capabilities

General-purpose robots must handle tasks, objects and environments they were never trained on, yet zero-shot generalization remains one of the least solved problems in physical AI. In digital AI, zero-shot capability grew with the volume and, above all, the diversity of training data, much of it sourced from the crowd. Robotics instead relies on paid collection by a few well-funded teams, which yields volume without diversity. We formulate zero-shot generalization in physical AI as coverage of the deployment distribution, and show that more samples from the same collectors shrink the unseen region only slowly, while each new contributor can extend it. Building on this and on zero-shot learning, we propose a shared latent space for robot actions that combines multi-modal alignment, generative augmentation, and noisy-label learning. As a scalable source of diverse data, we propose a community of developers building applications on one standard, approachable robot, Origami. We close with a testable hypothesis and an evaluation protocol relating real-robot zero-shot success to the diversity of contributed data.

Rafael Felix, Flavia Yanase

runibiLABS

Robot Policy Generalization Robot Learning Zero-Shot Robotic Manipulation Robotic Data Collection Zero-Shot Learning

Ask PDF

Sep 15, 2026

Embedding Physics Priors in Robot Learning: A Survey

The rapid progress of artificial intelligence is reshaping robotics and accelerating the adoption of learning-based approaches. While purely data-driven methods have achieved remarkable success in computer vision and natural language processing, robotics remains constrained by limited data, complex real-world interactions, and the need for reliable operation. These challenges have motivated the exploration of physics-embedded robot learning, which embeds physics priors into learning algorithms. By encoding the underlying physical laws and constraints, physics priors can complement limited data with robotics-specific inductive biases, potentially improving generalization, interpretability, and sample efficiency. However, the literature on physics-embedded robot learning remains fragmented across terminology, methodologies, and application domains, making it difficult to assess this growing body of work. This survey reviews physics-embedded robot learning across a broad range of physics priors, robotics applications, and machine learning models, from single-layer perceptrons to generative foundation models. We adopt a unified taxonomy that classifies existing approaches according to their physics embedding: physics-guided inputs, data, and representations; physics-encoded model architectures; and physics-informed training loss functions. Building on this taxonomy, we review methods for robot dynamics learning, trajectory planning, prediction, control, and estimation, together with the corresponding open-source software ecosystem. We identify key open challenges, and outline promising future research directions. Overall, we argue that physics priors provide a particularly relevant robotics-specific inductive bias, complementing rather than replacing data-driven learning, and paving the way toward more generalizable, data-efficient, and trustworthy robotic systems.

Mattia Piccinini, Lucas Schulze, Alice Plebe +10

Professorship of Autonomous Vehicle Systems, Technical University of Munich, 85748 Garching, Germany; Munich Institute of Robotics and Machine Intelligence (MIRMI). · Department of Computer Science, Technical University of Darmstadt, Germany. · Department of Industrial Engineering, University of Trento, 38123 Trento, Italy. +1

Robot Learning Physics-Informed ML Robotic Control Robot Dynamics Learning

Ask PDF

Feb 26, 2026

LeRobot: An Open-Source Library for End-to-End Robot Learning

Robotics is undergoing a significant transformation powered by advances in high-level control techniques based on machine learning, giving rise to the field of robot learning. Recent progress in robot learning has been accelerated by the increasing availability of affordable teleoperation systems, large-scale openly available datasets, and scalable learning-based methods. However, development in the field of robot learning is often slowed by fragmented, closed-source tools designed to only address specific sub-components within the robotics stack. In this paper, we present lerobot, an open-source library that integrates across the entire robot learning stack, from low-level middleware communication for motor controls to large-scale dataset collection, storage and streaming. The library is designed with a strong focus on real-world robotics, supporting accessible hardware platforms while remaining extensible to new embodiments. It also supports efficient implementations for various state-of-the-art robot learning algorithms from multiple prominent paradigms, as well as a generalized asynchronous inference stack. Unlike traditional pipelines which heavily rely on hand-crafted techniques, lerobot emphasizes scalable learning approaches that improve directly with more data and compute. Designed for accessibility, scalability, and openness, lerobot lowers the barrier to entry for researchers and practitioners to robotics while providing a platform for reproducible, state-of-the-art robot learning.

Remi Cadene, Simon Aliberts, Francesco Capuano +14

Hugging Face · University of Oxford

Robot Learning Robotic Control

Ask PDF

Abstract

Robot Learning

Explore similar work

Robot Generalization Beyond Volume: The Diversity Path via Zero-Shot Capabilities

Embedding Physics Priors in Robot Learning: A Survey

LeRobot: An Open-Source Library for End-to-End Robot Learning