available projects

2025

1. 

   Sampling-based MPC and Diffusion-based Online Policy Distillation for Agile Control of Hybrid Mobile Robots

   Jin Cheng and Stelian Coros

   bipedal-locomotion, trajectory-optimization, model-predictive-control, policy-distillation, 2025

   Abs PDF Website

   Hybrid mobile robots that combine wheels and legs offer versatile locomotion capabilities but pose significant control challenges due to their complex dynamics and multi-modal interactions. Traditional reinforcement learning approaches can train task-specific policies but often suffer from the sim-to-real gap and lack mechanisms for real-time adaptation. Model-based methods such as model predictive control (MPC) provide robust, high-quality actions but typically require accurate dynamics models and state estimation, making them difficult to deploy directly on hardware. This project proposes a hybrid control-learning pipeline that leverages state-of-the-art sampling-based MPC for real-time action optimization while simultaneously performing online policy distillation with diffusion models. Sampling-based MPC generates reliable control signals by exploiting fast, parallelized simulation, and these signals are distilled into neural policies that capture the multi-modality of the optimizer’s output. The resulting policies combine the robustness of model-based optimization with the efficiency of learned inference, enabling agile, deployable controllers for hybrid wheeled-legged robots.

2. 

   Active Perception for Ball Catching with a Quadrupedal Manipulator

   Jin Cheng, Tianxu An, and Boyang Sun

   Loco-manipulation, Active Perception, Reinforcement Learning, 2025

   Abs PDF Website

   Catching fast-moving objects presents a significant challenge for mobile manipulators due to the need for tightly coupled perception and control under dynamic conditions. This project aims to develop a unified visuomotor control policy for a quadrupedal robot equipped with an arm-mounted RGB-D camera, enabling it to perceive, track, and intercept a thrown ball using only onboard sensing. By mounting the camera on the end-effector, the robot gains the ability to adjust its viewpoint in real time, mimicking human strategies of active perception. The control policy will be trained entirely in simulation with randomized ball trajectories and perception noise models to support robust, zero-shot deployment in the real world. The resulting system will advance research at the intersection of active perception, reinforcement learning, and dynamic whole-body manipulation.

3. 

   Reinforcement learning for robotic 3D bin packing

   Valentin Hartmann

   Reinforcement Learning, Manipulation, 2025

   Abs PDF Website

   Many industrial assembly tasks require high accuracy when manipulating parts, e.g., when inserting an object somewhere, such as a box that we would like to put in a bin. Here, the tolerances that are required can vary wildly. While industrial manipulators are generally accurate in a controlled setting, as soon as the environment is not as structured as we would like it to be, we need to leverage feedback policies to correct for inaccuracies in estimation, and unexpected/unmodeled effects. We are interested robotic bin packing, i.e., enabling a robot to pick boxes/parcels, and placing them in a box, where other previously placed object possibly obstruct the placement.