Abstract
Humanoid robot control often struggles to balance the natural appearance of imitation learning with the physical consistency of model-based methods. This paper introduces a hybrid architecture that embeds kinematic constraints directly into a neural network's loss function and incorporates a steering correction block during inference. Evaluated on the ergoCub humanoid, the proposed method significantly reduces foot sliding and eliminates directional drift, providing a robust framework for real-world deployment.
Introduction
Generating smooth, human-like trajectories for humanoid robots is a significant challenge in robotics. Traditional model-based methods, such as those relying on inverted pendulum models, offer stability but often result in stiff, "robotic" movements. In contrast, model-free approaches like Imitation Learning (IL) or Behavior Cloning (BC) can produce natural-looking motion by mimicking human data. However, these methods often lack an inherent understanding of physics, leading to several critical issues:
- Foot Sliding: The robot's "planted" foot may move on the ground because the network has not learned the hard constraint of zero contact velocity.
- Directional Drift: Small biases in training data or network approximations accumulate, causing the robot to veer off its intended path.
- Data Scarcity: Learning high-dimensional humanoid motion purely from observation requires massive datasets.
This research proposes a two-pronged strategy to bridge the gap between learning-based motion generation and physical reality.
Methodology
The authors propose a framework that combines a physics-informed training process with a control-informed deployment strategy.
Physics-Informed Loss (PI Loss)
The core architecture utilizes a Mode-Adaptive Neural Network (MANN), a "Mixture of Experts" model that predicts the robot's next state based on its current state and user input. While standard training uses Mean Squared Error (MSE) to match human data, this work introduces a Physics-Informed (PI) loss term ($L_B$). By utilizing the robot's Jacobian matrix—which maps joint velocities to foot velocities—the loss function penalizes predicted motions where the support foot has a non-zero velocity. This forces the network to prioritize kinematically feasible motions.
Control-Informed Steering
To address directional drift, a Proportional-Integral (PI) correction block is implemented at the inference stage. This block acts as a "steering" mechanism, comparing the network's predicted trajectory to the user's desired goal (position and orientation) and applying corrective terms to the velocities before they reach the low-level controllers.
Data and Implementation
The model was trained on 30 minutes of multi-directional human walking data captured via XSens motion capture suits. This data was retargeted to the ergoCub humanoid robot, a 160 cm tall, 56 kg platform with 26 degrees of freedom. Data augmentation through sagittal mirroring was used to ensure motion symmetry.
Results and Discussion
The integration of physical priors and steering correction yielded significant improvements in trajectory stability and accuracy.
Impact of Physics-Informed Loss
The researchers found that the weight ($w$) of the PI loss term is critical. If $w$ is too high, the model may "cheat" by satisfying the constraint in non-natural ways (e.g., rotating on a single point). A "sweet spot" of $w=10$ was identified, which effectively balanced physical constraints with natural motion. As shown in the results, the physics-informed approach slashed the support foot's angular velocity error (RMSE) from approximately 17 rad/s to 3 rad/s.
Drift Elimination and Robustness
In long-duration walking tests (10 seconds), the baseline model exhibited significant lateral and yaw drift. The addition of the steering block virtually eliminated this drift, keeping the robot on its intended path. Furthermore, the system demonstrated robustness across different control backends, including offline Divergent Component of Motion (DCM) and online Model Predictive Control (MPC) controllers.
Conclusion
This work demonstrates that embedding physical constraints into the learning process and adding modular control blocks can significantly enhance the performance of imitation learning for humanoid robots. Future work will focus on incorporating full dynamics—such as ground reaction forces and centroidal momentum—to improve stability on uneven terrain and automating contact switching within the network architecture.
Summary updated