AI & Robotics Engineer

The core of my M.S. thesis, framed as an alignment problem for embodied agents. A robot can replicate a human's joint configuration exactly and still lose what the motion means: a retargeted "drinking" gesture that never reaches the mouth is kinematically correct but semantically wrong. I added a post-refinement stage where an agent learns per-joint adjustments for four key joints (shoulder, arm, forearm, wrist) via RLHF: a reward model trained on human-preference comparisons supplies the reward, and PPO optimizes the agent to recover the semantic intent that pure kinematic retargeting discards. In a human survey, refined motions were consistently preferred over the un-refined baseline.

RLHF aligns motion to human intent, but human feedback is slow, costly, and doesn't scale — the bottleneck of my thesis. This prototype swaps the human evaluator for a learned one — RLAIF applied to robotics: a CLIP-style reward model that automatically scores how well a robot's motion matches the semantic intent of a natural-language description ("drinking motion", etc.), pairing a Transformer motion encoder with a fine-tuned LLaMA language encoder in a shared embedding space where cosine similarity becomes the reward. It turns a costly human-feedback reward model into a scalable, automatic one — the direction my thesis itself flagged as its missing piece.