Fast Humanoid Loco-Manipulation via Flow Matching
Overview
A course project for MIT 6.4210 (Robotic Manipulation) comparing DDPM and Flow Matching for humanoid whole-body control via trajectory synthesis.
Key Results
- Flow Matching achieves 820 vs. 280 survival steps at 5 neural function evaluations (NFE), dramatically outperforming DDPM at low compute budgets
- Zero-shot loco-manipulation from walking-only training data via test-time classifier guidance — no manipulation demonstrations needed
- Demonstrates that FM’s straight-path interpolation is fundamentally better suited for fast trajectory generation than DDPM’s curved diffusion paths