DynaGuide: Steering Diffusion Polices with Active Dynamic Guidance

Maximilian Du, Shuran Song

Neural Information Processing Systems (NeurIPS) 2025

Website  /  Paper  /  Code

Robot policies are getting more complicated, which also means that tuning them also becomes tedious. In DynaGuide, we propose a novel way of steering policies by using a dynamics model to influence the action inference process of a robot policy. DynaGuide works on any diffusion policy, including off-the-shelf real robot policies.

To Err is Robotic: Rapid Value-Based Trial-and-Error during Deployment

Maximilian Du, Sasha Khazatsky, Tobias Gerstenberg, Chelsea Finn

Website  /  Paper

When deploying robot policies, we need to know when we've made mistakes so we can recover from them and try something else. In this paper, we propose a mistake detection approach that uses the Bellman error of a learned value function. This simple trick boosts success rates of trained policies, all without modifying the base policy.

Behavior Retrieval: Few-Shot Imitation Learning by Querying Unlabeled Datasets

Maximilian Du, Suraj Nair, Dorsa Sadigh, Chelsea Finn

Robotics: Science and Systems 2023

Website  /  Paper  /  Code

When training robots on large datasets, we need to select the right data to improve its performance without confusing the robot. In this work, we propose a simple data filtering approach using distances in a latent representation. This data filter has large impacts on robot performance in simulation and on a real setup.

Play it by Ear: Learning Skills amidst Occlusion through Audio-Visual Imitation Learning

Maximilian Du*, Olivia Lee*, Suraj Nair, Chelsea Finn

Robotics: Science and Systems 2022

Website  /  Paper  /  Code

Operating in the real world, sound cues can tell us information even when vision cues fail, like rooting around a dark bag. In this work, we add the sound modality to a robot by adding a microphone to the gripper. We show that it is able to locate and extract hidden keys from a bag.