Robotics Student Projects in CRL

Can we optimize a design so that a robot can assemble it more easily? This project explores inverse design for robotic assembly using task and motion planning (TAMP), coupling design optimization with planning algorithms that lays the foundation for designing robot-friendly structures.

Hybrid mobile robots that integrate wheels and legs present a promising locomotion paradigm, combining the speed and efficiency of wheeled motion with the adaptability of legged systems. However, controlling such platforms remains challenging due to their underactuated and hybrid dynamics. This project explores a control-learning framework that leverages trajectory optimization (TO) or model predictive control (MPC) to generate high-quality motion plans while concurrently distilling a lightweight neural policy through supervised learning. The approach supports both sampling-based and gradient-based optimization methods, depending on simulator capabilities, and emphasizes real-world learning by collecting rollout data directly from hardware. The resulting policy enables efficient and robust control suitable for deployment under real-time constraints. Development and evaluation will be conducted using the LIMX Dynamics TRON1 robot, a bipedal wheeled platform with 4 DoF per leg, providing a rich testbed for advancing agile and intelligent control in hybrid robotic systems.

Robotic motion planning typically assumes static, explicitly defined environments, limiting a robot’s operational capabilities. This project explores how robots can actively modify their surroundings to dynamically extend their workspace. By integrating physics-based simulation with sampling-based motion planning techniques, we aim to enable flexible mobile manipulation in changing environments.

Manipulating large, heavy objects without convenient grasp affordances poses a significant challenge for humanoid robots. This project develops a demonstration-guided reinforcement learning framework enabling a humanoid robot to strategically lift unwieldy objects through coordinated full-body motion.

Robotic task and motion planning (TAMP) for deformable objects typically relies on predefined physics models, limiting the robot’s adaptability in real-world scenarios. This project explores how robots can autonomously construct physics-informed digital twins directly from videos and integrate the learned physics model into a TAMP framework to achieve "on-demand" manipulation of deformable objects.

This project aims to enable agile and adaptive behaviors in real-world quadruped robots through zero-shot reinforcement learning (RL). Traditional RL methods often rely on fixed reward functions, limiting their generalization to novel tasks. In contrast, the forward-backward (FB) framework learns representations that allow policies to generalize across a wide range of goals without retraining. We propose to implement and evaluate the FB method on the Unitree Go2 quadruped robot, benchmarking it against alternative zero-shot RL approaches. By leveraging recent advances in exploration strategies and representation learning, this work seeks to demonstrate zero-shot capabilities on a physical robot for tasks such as novel gait synthesis, pose reaching, and motion imitation. Success in this endeavor could significantly broaden the applicability of RL to real-world robotics, laying the groundwork for flexible, reward-agnostic robotic systems.