Ο€β‚€.β‚… Piper stack_blocks_three

Private OpenPI checkpoint from full fine-tuning of pi05_base on HorizonRobotics/Real-World-Dataset v0.2. Independent single-task run (not multi-task). Optimizer train_state is not included.

OpenPI config pi05_piper_stack_blocks_three
Experiment stack_blocks_three_ft10k
Recommended step 8000
Train steps 10000
Offline val FM loss 0.01254
Train / val episodes 180 / 20
Prompt Stack the three blocks on top of each other.
FSDP / batch / EMA 2 / 8 / 0.99
LR cosine warmup 1000, peak 2.5e-5, decay 30000 β†’ 2.5e-6

Layout

  • params/: Orbax PyTree weights (OpenPI weight_loader / policy serve)
  • assets/: train-split norm_stats.json
  • config.json: run metadata

Load with OpenPI by pointing --policy.config pi05_piper_stack_blocks_three and the checkpoint directory at this repo (the folder that contains params/).

Downloads last month
7
Video Preview
loading