Efficient Inference for Embodied Foundation Models
Efficient Machine Learning · Robotics
Selected project · Nov 2025 – Mar 2026
Advisor: Dr. Jiachen Liu
Accelerating world action models for real-time robotic decision-making.
My contributions
- Analyzed inference efficiency in world action models (WAMs) for robotic policies.
- Implemented cross-attention KV caching and token compression to accelerate inference.
- Explored dynamic-precision quantization and speculative decoding for more efficient agent decisions.