Active projects

Current research led by members of the Berkeley Intelligent Control Lab.

Multi-Agent Systems

Decentralized Multi-Robot Collaboration via Higher Order Reasoning

This project is focused on multi-robot collaboration in decentralized settings where each robot only has access to its own observations, so successful coordination requires robots to infer what their teammates are doing or know and signal their own...

Student lead: Ethan

Learning for Control

Trajectory Guidance using Diffusion as a Candidate Generator

This project focuses on using a diffusion model as a candidate generator to produce many different possible solutions. Then, with a learned model, we can evaluate those possible solutions to select the best option. This will also give...

Student lead: Ethan

Learning for Control

Continual Learning in VLAs

This project addresses continual learning in Vision-Language-Action (VLA) models to reduce catastrophic forgetting. Instead of using replay buffers or co-training, it applies constrained optimization to limit parameter drift away from weights optimal for prior tasks.

Student lead: Aayushi

Other

Learning from Only Positive Demos

In many real-world deployment settings, demonstrations of negatives (failures) are scarce or entirely unavailable. This project asks: what is the best way of learning from only positive (successful) demonstrations without limiting the learning to behavior cloning?

Student lead: Lasse

Other

Whole Body Tool Use

Humans are incredibly good at leveraging their entire body when interacting with the world and tools around them: opening doors with a foot while holding a tray or package, using an elbow to push a button when their...

Student lead: Lasse

Other

3D-Predictive Object Detection for False Positive Robustness

Current open-vocabulary object detection performs well in clean scenarios. However, in actual use cases, most images are from suboptimal viewpoints, contain many occlusions, or do not contain relevant information at all. Even when aggregating information from multiple views,...

Student lead: Marfred