Unnat Jain

Results show that GLIDE can support policy learning when direct demonstrations are infeasible, and discovers emergent guardrails that go beyond domain-expert hardcoded ones, improving data collection over naive VR teleoperation and domain-expert hardcoded guardrails.

Shifting the training focus from low-level actions to a unified representation can dramatically enhance the efficiency and robustness of robot learning systems.

DeVA achieves faster convergence and superior performance in robot manipulation tasks by decoupling video and action predictions while integrating physical guidance.

Achieving up to 60% success in robot manipulation by aligning language and action predictions without sacrificing pretrained visual representations reveals a breakthrough in VLA policy training.

A novel anthropomorphic hand design achieves robust and versatile manipulation by strategically combining hard and soft materials, offering improved strength, endurance, and grasp success at a low cost.