Introduction
Modern robotics often relies on hand-tuned controllers designed for specific tasks. However, as robots enter real-world environments with varying terrains and physical constraints, fixed control strategies often prove insufficient. This research explores how a mobile robot can autonomously learn the "physics" of its own body to perform complex tasks, such as using its arms to maintain stability on an incline.
Problem Statement: The Need for Adaptive Control
Traditional controllers are typically optimized for a specific configuration. When an embodied agent encounters a new context—such as a steep slope or a slippery surface—a static controller may fail. The challenge lies in model learning for whole-body control: enabling a robot to discover the relationship between its various effectors (e.g., arms) and its overall state (e.g., velocity and balance). The goal is to allow the robot to utilize any available degree of freedom to achieve a global objective, even if that part of the body wasn't originally intended for that specific function.
Methodology
The researchers proposed a two-phase approach: learning a forward dynamics model and then applying it for control.
1. Model Learning via Lasso and BIC
The robot represents its dynamics using a linear-Gaussian basis function model. To prevent overfitting and eliminate the need for manual parameter tuning, the Lasso algorithm ($l_1$ regularization) is employed for basis selection. Lasso effectively forces the weights of irrelevant features to zero, identifying a parsimonious model. The Bayesian Information Criterion (BIC) is then used to automatically select the optimal model from the candidates, making the learning process largely parameter-free.
2. Jacobian-Based Control
Once the forward model is learned, the robot calculates a Jacobian matrix. This matrix maps small changes in action (arm velocity) to changes in state (wheel velocity). By utilizing the pseudoinverse of this Jacobian, the robot can invert the learned physics. It defines an objective function—such as minimizing wheel velocity—and calculates the necessary arm movements to descend the error gradient.
Experimental Setup
The experiments were conducted using the uBot-5, a dynamically stable balancing robot.
The robot collected four minutes of self-generated exploratory data by performing random arm movements while balancing. The state space consisted of 10 continuous dimensions, with a 2-dimensional action space for shoulder velocities. From 90 candidate basis functions, the Lasso/BIC process identified approximately 60 essential physical relationships.
Key Results
The effectiveness of the learned model was tested by placing the robot on a 10-degree incline.
As shown in the performance data, a standard tuned LQR controller allowed the robot to drift down the ramp. In contrast, the learned controller enabled the robot to "discover" that extending its arms outward shifted its center of mass, creating a compensatory moment that counteracted gravity. The robot reached a steady dynamic equilibrium after rolling back only 25 cm, effectively holding its ground through whole-body coordination.
Conclusion
By treating effectors as general-purpose tools rather than task-specific components, this approach allows robots to adapt to novel physical challenges. Future work aims to improve exploration efficiency in higher-dimensional spaces and demonstrate the reuse of these learned models across a wider variety of tasks.
Summary updated