Chi-Pin Huang

Enables fast, accurate visual planning with a lightweight drafter that generates an entire visual primitive, such as a bounding box or trajectory, in one forward pass.

Learns from both human and robot videos by reasoning in a shared motion space and imagining future world states, rewarded by what actually happens next.

Guides robots to actively probe hidden physical properties, such as mass and stiffness, for manipulation when vision alone is insufficient.

Compresses textual reasoning into compact, verbalizable latent plans for faster vision-language-action inference.

Connects embodied reasoning to robot actions through reinforced visual latent planning, supporting long-horizon tasks and self-correction.

Combines multiple subject identities and reference motions to generate customized videos with consistent appearances and interactions.

Uses lightweight erasers to suppress target concepts in diffusion models while preserving unrelated concepts and resisting adversarial prompts.