PI 0.5 PiperX handle-grasp checkpoint (step 10,000)
This repository contains the complete step-10,000 training checkpoint from the PiperX handle-grasp fine-tuning run. It includes both model parameters and optimizer/train state under checkpoint-10000/.
- Base policy: PI 0.5
- Robot: Piper X
- Task instruction:
grasp the handle - Camera inputs: rear and wrist
- Dataset: 250 episodes / 197,545 frames from the private dataset repository ishitag925/handle-grasp-piperx
- Training: standard PI 0.5 loss, batch size 4, 50,000-step run configured with checkpoints every 10,000 steps
- Evaluation at this checkpoint: 0/5 successful grasps on five training-distribution poses; each rollout timed out at 45 simulated seconds using bilateral finger contact as the success criterion
The training run continues independently. This is a snapshot; it does not include the later checkpoints. To resume or evaluate it, use the matching Piper training code and config from the experiment.
- Downloads last month
- -