RoboPapers

Ep#6: FP3: A 3D Foundation Policy for Robotic Manipulation

Apr 24, 2025
Ask episode
AI Snips
Chapters
Transcript
Episode notes
ANECDOTE

Gao Yang's Robotics Journey

  • Gao Yang transitioned from computer vision to robotics due to interest in combining vision with motion capabilities.
  • He is now an assistant professor and founder of Spirit.ai, focusing on bimanual general-purpose robots.
INSIGHT

3D Inputs Boost Success

  • Using 3D point cloud inputs instead of 2D images greatly improves policy success rates.
  • FP3 achieves about 80% success in wild scenarios compared to near zero for 2D-based baselines.
ADVICE

Pre-train Then Fine Tune

  • Pre-train policies on large-scale, diverse datasets like Droid containing 3D information.
  • Use LoRa fine tuning with a small number of demos to adapt to specific tasks efficiently.
Get the Snipd Podcast app to discover more snips from this episode
Get the app