jiankimr/pi05_franka_stack_cube_50ep_seed42_ep20

pi0.5 · Franka · stack the orange cube on the blue cube · seed 42 · epoch 20

Full fine-tune of pi05_base on the real-robot dataset jiankimr/stack_cube_50ep_lerobot (revision b082e72f, 50 episodes, 9,368 frames, 15 fps).

Other checkpoints of the same run: ep5, ep7, ep10, ep12, ep15, ep18, ep70.

epoch	20 (1 epoch = 292 steps at batch 32)
Orbax step	5840
batch size	32, single H100, FSDP 1
LR schedule	cosine, warmup 1,000, peak 2.5e-5, decay_steps 30,000, final 2.5e-6
EMA decay	0.99
seed	42
training loss (step 5800)	0.0086

Training ran to 70 epochs (20,440 steps); this repository is the step-5840 snapshot. Trained in one uninterrupted run (job #263).

Contract

Trained with the rl-max/openpi_franka at d79defa03d339012d6a7c2c01c6c6f8c74e281e7 fork, config pi05_cube_stack_orange_on_blue:

• Pi0Config(pi05=True, action_dim=32, action_horizon=16)
• LeRobotCustomDROIDDataConfig — model sees exterior_image_1_left + wrist_image_left
• state: 14-D stored, [xyz, rpy, 7 joint positions, gripper]; the model uses [xyzrpy, gripper]
• action: 7-D commanded cartesian_delta, [dx, dy, dz, droll, dpitch, dyaw, absolute_gripper]
• norm stats: assets/stack_cube_50ep_lerobot/norm_stats.json, computed from this dataset

The evaluation instruction must match training exactly:

stack the orange cube on the blue cube

Download and serve

The repository root already contains params/ and assets/, so it can be passed straight to the server. Downloading with local_dir keeps the files out of the Hugging Face cache; delete the directory when finished.

huggingface-cli download jiankimr/pi05_franka_stack_cube_50ep_seed42_ep20 --local-dir ./checkpoints/ep20

XLA_FLAGS="--xla_gpu_enable_triton_gemm=false" \
uv run python serve_checkpoint.py --checkpoint-dir ./checkpoints/ep20 --port 8000

train_state/ is not included; these checkpoints are for inference only.