pi05_allslot_st3_s24000

• Model type: pi05 (paligemma_variant gemma_2b, action_expert gemma_300m, bfloat16, chunk_size 50, n_action_steps 50, num_inference_steps 10).
• Initialization: fine-tuned directly from lerobot/pi05_base (pretrained_path in config.json).
• Training data: 78-episode four-slot scripted-chain demonstrations of the Task 2 thermal-pad lay (local/task2_allslot_st3_s29), recorded with the benchmark's own LeRobot recorder (30 fps, eval_split 0.05, pyav backend).
• Steps: checkpoint at 24 000 optimization steps.
• Action space: 20-dim whole-body action row — base twist (vx, vy, wz), left + right FR3 7-DOF absolute joint targets, left + right gripper open-fraction, spine.height.target (see action_feature_names in config.json).
• Observation: 32-dim state + three 224×224 RGB cameras — observation.images.base_0_rgb, observation.images.left_wrist_0_rgb, observation.images.right_wrist_0_rgb. STATE/ACTION normalization: QUANTILES (pre/post processor pipelines are shipped beside the weights and must be loaded with them).
• Intended use: EBiM Task 2 evaluation. Load with LeRobot ≥0.6 as a pi05 policy — point the deployment config's vla.checkpoint at this directory (see the submission repo's "Fine-tuning on this task's own demonstrations" / LeRobotBackend, which loads the processor pipelines saved beside the weights so inference normalization is identical to training).

Files