fbsh96/rebot_act_flipbread_44eps_mi300x_b16_10000steps

reBot ACT Flip-Bread 44-Episode MI300X Checkpoint

This is a LeRobot ACT policy checkpoint trained on the merged 44-episode reBot/B601 flip-bread dataset:

• Lisette1231/20260425_flipbreadtopot1
• Lisette1231/20260425_flipbreadtopot2
• Lisette1231/20260425_flipbreadtopot3
• Lisette1231/20260425_flipbreadtopot4_newway
• Lisette1231/20260425_flipbreadtopot5_newway

Model

• policy type: act
• robot type: seeed_b601_dm_follower
• state/action dimension: 7D
• image inputs: observation.images.front, observation.images.wrist
• action horizon: 50
• training host: AMD Instinct MI300X VF / ROCm
• training steps: 10000
• batch size: 16
• final training loss: about 0.124

Validation

The checkpoint was reloaded successfully and validated on MI300X with synthetic reBot observations:

• predict_action_chunk output: (1, 50, 7)
• DRTC loopback returned action chunk: (50, 7)
• DRTC loopback after warmup: about 72 ms
• server observation-to-action-send: about 48.9 ms

Safety Note

This is a validation / offline imitation checkpoint. It has not been proven safe for autonomous real-robot closed-loop execution. Use logging-only replay, action clipping, rate limits, joint limits, and an emergency stop path before any actuator execution.