PhyFilter: Physics-Based Feedback Lets Robots Generalize Without Massive Data

Background: The Data-Scaling Bottleneck in Robot Learning

Large language models have achieved remarkable progress through massive data, compute, and model scaling. However, transferring this paradigm to robotics faces a fundamental obstacle: robot training data must be collected in the physical world. Human teleoperation is slow, expensive, and difficult to scale to LLM-level datasets (hundreds of billions of tokens); even Tesla's large-scale efforts remain bottlenecked by collection throughput. Simulation engines (Isaac Lab, MuJoCo) and video generation models (Sora, Cosmos) can synthesize data, but they still suffer from sim-to-real gaps, high computational cost, and a lack of standardized cross-platform data formats and hardware interfaces.

The authors turn to biology for inspiration. Evidence shows that biological generalization does not rely solely on scale. Innate or acquired physical structure itself promotes adaptation to unseen environments: cortical motor responses are governed by feedforward signals, feedback loops, and prior-driven dynamics that evolve to minimize a global energy function; zebrafish larvae integrate innate physical integration structures with real-time state feedback into an updating cerebellar internal model to respond effectively to unpredictable visual disturbances.

This raises the central question: unlike digital-world LLMs, real robots already possess well-structured physical priors and continuous real-time feedback from environmental interaction. Can these "ready-made" physical structures be exploited to improve the generalization of robot learning?

Core Method: Treating Learning Residuals as a Filterable Signal

PhyFilter starts from a quantity overlooked by prior work — the learning residual , defined as the difference between the true mapping and the learned model. This residual arises from sensor noise, external disturbances, or unmodeled interactions in unseen scenarios and is the root cause of generalization failure.

Ideally, if the learning residual were known in real time, the model output could be corrected immediately. However, during deployment no ground-truth labels are available, so the residual cannot be obtained instantly. Biological systems likewise do not "instantly understand" unfamiliar environments but undergo an adaptation process.

Inspired by the low-pass filtering characteristics ubiquitous in human visual and auditory perception, the team formulates the objective as a high-order low-pass filter. If this filter form can be realized, the low-frequency features of the learning residual can be convergently captured and compensated in time; the convergence process of the low-pass filter itself can be viewed as the robot's self-adaptation in unseen scenes.

The key technical difficulty is that the learning residual is not measurable. The team's solution introduces the robot's own real-time state feedback and prior differential-equation structure, equivalently rewriting the original form — which required high-order derivatives and the learning residual — into an implementation that depends only on measurable signals, thereby completely avoiding direct high-order differentiation of the nominal model.

The resulting PhyFilter has three key properties:

Plug-and-play. Unlike previous physics-informed methods that require retraining the modified network, PhyFilter is a model-agnostic lightweight module that can be deployed directly on a pre-trained model.

Degradable and extensible. Under a first-order filter setting, the implementation degrades to a pure feedback disturbance observer; higher-order forms handle complex residuals beyond first-order dynamics.

Auto-tunable parameters. Filter parameters traditionally require tedious manual pole placement, which becomes unintuitive at higher orders. The team reinterprets filter parameters as control inputs of a dynamical system, transforms tuning into an optimal control problem, and solves gradients analytically using Lagrange multipliers with the adjoint method (reverse-mode differentiation), optimized with Adam, mini-batches, and early stopping. Experiments show auto-learned parameters match carefully hand-tuned ones.

Computationally, PhyFilter runs at 500 Hz on an STM32F765 microcontroller in drone flight and aerial manipulator experiments, demonstrating suitability for resource-constrained lightweight platforms.

Evaluation: Four Real Robot Systems

The team validates PhyFilter on four representative scenarios: quadruped locomotion (RL), drone agile flight (SL), aerial manipulation (SL), and acceleration estimation (SL), covering both mainstream robot learning paradigms.

Quadruped Locomotion: Flat-Ground Simulation to Four Real Terrains

To highlight generalization, the simulation training environment contained only flat ground with standard parameter randomization. Across five random seeds, policies augmented with PhyFilter achieved consistently higher maximum training returns and more coordinated gaits than baselines, indicating a robust low-level controller significantly improves upper-level RL training efficiency.

In generalization tests, the robot was commanded to run at 2.83 m/s (outside training distribution) while carrying 15%–60% of its body weight (11.86 kg), up to 137.33% beyond the training load limit. On five unseen terrains — stairs, horizontal rough terrain, sloped rough terrain, wavy ground, and gravel — PhyFilter success rates were 70%, 80%, 50%, 80%, 40% versus baseline 0%, 0%, 10%, 20%, 0% and RL2AC (adaptive control) 0%, 20%, 10%, 50%, 0%.

Real-world deployment used zero post-deployment tuning. The baseline policy quickly became unstable and toppled on sand and gravel, while PhyFilter-guided policies walked stably on all terrains. Notably, even when the correction term was disabled at deployment, the policy still transferred successfully — evidence that PhyFilter shapes a fundamentally better policy, not merely a runtime patch.

Drone Agile Flight: Plug-and-Play Generalization Boost

In the supervised learning setting, PhyFilter was stacked on top of SEER-I, which combines offline basis-function learning with online adaptive control to capture mass uncertainty but degrades when unknown disturbances like wind are introduced.

On a 2 m/s circular trajectory tracking task, adding PhyFilter reduced mean absolute tracking error by 30.22% vs. SEER-I and 50.17% vs. baseline. Uncertainty estimation comparison showed SEER-I captures only mass uncertainty, while PhyFilter simultaneously identifies wind-induced disturbances and inherent dynamical uncertainties.

Aerial Manipulator: Centimeter-Level Pick-and-Place Under 5 m/s Wind

The aerial manipulator must achieve high-precision end-effector control under strongly coupled dynamics and external disturbances. An emergency-rescue-style pick-and-place task required retrieving medicine from a tripod and placing it in a designated zone with a maximum allowed tracking error of 2.5 cm. A 380 W fan generated up to 5 m/s wind, plus an additional 0.3 kg unmodeled payload.

Both baseline and SEER-I failed due to unseen wind and mass uncertainty; only PhyFilter completed all four stages, improving tracking error mean and variance by 23.99% and 55.04% respectively over baseline.

Acceleration Estimation: Reliable Neural Differentiator Under Distribution Shift

For lightweight platforms like drones lacking high-precision sensors, acceleration estimation is challenging. A two-hidden-layer neural network was trained on real flight data to map historical velocity to acceleration; training velocities ranged 0–2 m/s, test velocities 0–3 m/s. The pure learning model exhibited clear generalization failure on the test set, while PhyFilter — incorporating kinematic structure knowledge — achieved the highest accuracy, outperforming carefully tuned classical baselines: Robust Differentiator (RD), Tracking Differentiator (TD), and Discrete Differentiator (DD). This advantage held consistently across sampling intervals of 0.002–0.018 s.

Unified Generalization and Precision Without Excessive Domain Randomization

Viewing the four cases together, the conclusion is clear: mainstream paradigms bridge sim-to-real gaps via domain randomization — essentially seeking average performance across many random environments, trading precision for robustness, akin to classic robust control. PhyFilter, by contrast, unifies generalization and precision in a single framework without excessive domain randomization. When the robot enters out-of-distribution scenes, PhyFilter activates to enhance generalization; in familiar regimes where residuals are small, its intervention adaptively decays, preserving original precision. This also means training can save the substantial engineering effort spent crafting effective randomization parameters.

Limitations and Future Directions

The authors identify three limitations that define future work:

PhyFilter currently only enhances the learning output . A smarter approach would feed knowledge gained during filtering back to refine the original network's parameters or structure; continual learning may be a viable path.

The auto-tuning algorithm, while reliable, requires a full forward trajectory rollout per iteration, limiting efficiency on large-scale data; truncated or parallelized rollouts are promising improvements.

The learning residual may mix neural prediction errors with other uncertainties (e.g., modeling mismatch in the quadruped case). Disentangling these components — possibly via meta-learning — remains an open problem.

Overall, PhyFilter offers a complementary direction to "data scaling": robots are not image generators lacking physical priors; they inherently possess well-structured dynamics models and continuous real-time feedback. Leveraging these existing physical structures may be closer to the essence of biological generalization than endlessly expanding datasets.

References and Resources

Paper: https://www.nature.com/articles/s44182-026-00114-y (DOI: 10.1038/s44182-026-00114-y)

Project Website: https://scoardyy.github.io/PhyFilter/ Code Repository:

https://github.com/JIAjindou/PhyFilter