Research Goals
Our mission is to develop embodied intelligent agents capable of solving tasks that benefit humans. We accomplish this primarily through simulation and sim2real transfer, with a focus on minimizing—though not eliminating—real world robot data.
Cross-Embodiment Policies
We’re building towards one policy to rule them all: a single unified policy for planning, perception, and control across multiple robotic platforms, including:
- Humanoids
- Bimanual mobile platforms
- Quadrupeds
Modular Architecture
Our approach decouples the three critical components of embodied intelligence:
- Low-Level Control - Developing a low-level motion control policy
- Perception - Vision models coupled with proprioception
- Planning - Language model-driven task planning
Learning Pipeline
We’re developing new architectures that enable end-to-end learning, building upon the state of the art in robot learning:
Webscale Pretraining → Robot Data Pretraining → Task Fine-tuning → RL
Current Research Projects
1. Full-Body Humanoid Teleoperation
Building a full-body humanoid teleoperation system using Meta Quest 3 for high-quality dataset collection. This creates the foundation for learning human-like manipulation and locomotion behaviors.
2. Cross-Embodied Motion Control Policy
Developing a cross-embodied adversarial skill embedding style low-level motion control policy for robust sim2real transfer. This project works in tandem with the teleoperation system to enable skills learned in simulation to transfer seamlessly to real-world robots across different embodiments.
3. Vision-Language-Action Model
Creating a vision/video language action model with an action expert component that produces embeddings to drive the low-level motion control policy. This bridges high-level perception and language understanding with low-level motor control.