DiscoDemo-Stage3_SFT-FMB_Round-alpha0

A visuomotor π0.5 policy for the FMB-Round task (insert a cylindrical peg into its hole on the FMB board), fine-tuned on demonstrations generated by the P-RFCL baseline generator (α = 0: the same reverse-curriculum RL without the diversity reward) of DiscoDemo. It acts from camera images and joint states only, and transfers to the real Franka FR3 robot without real-world fine-tuning.

Stage	Repository
Generator (RL policy)	DAVIAN-Robotics/DiscoDemo-Stage1_RL-FMB_Round-alpha0
Generated dataset	DAVIAN-Robotics/DiscoDemo-Stage2_GenData-FMB_Round-alpha0
Imitation policy (π0.5)	DAVIAN-Robotics/DiscoDemo-Stage3_SFT-FMB_Round-alpha0 (this repository)

Training

Base model	DAVIAN-Robotics/pi05_droid_jointpos (π0.5 DROID, joint-position actions)
Data	DAVIAN-Robotics/DiscoDemo-Stage2_GenData-FMB_Round-alpha0
Steps	20,000
Batch size	16
Action chunk	15 (relative to the current joint state)
Cameras	over-the-shoulder and wrist
Seed	1000

Usage

The policy is a LeRobot π0.5 checkpoint. Its pre- and post-processing pipeline uses DiscoDemo processing steps (relative actions, state tokenization), so loading it requires the DiscoDemo code: https://github.com/DAVIAN-Robotics/DiscoDemo.

License

Apache-2.0, following the base model (openpi).

Citation

@article{park2026discodemo,
  title   = {DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning},
  author  = {Park, Minho and Kim, Kinam and Kim, Donghu and Lee, Byungkun and Hwang, Dongyoon and Shin, Yongjae and Hyung, Junha and Lee, Hojoon and Choo, Jaegul},
  journal = {arXiv preprint},
  year    = {2026}
}