fbsh96/rebot_smolvla_flipbread_44eps_20260425_3000steps

reBot SmolVLA Flip Bread 44 Episodes

SmolVLA checkpoint fine-tuned for seeed_b601_dm_follower on merged LeRobot flip-bread-to-pot demonstrations collected on 2026-04-25.

Schema

• observation.state: 7D
• observation.images.front: (3, 480, 640)
• observation.images.wrist: (3, 480, 640)
• action: 7D
• action chunk: 50 x 7

Joint/action order:

shoulder_pan.pos
shoulder_lift.pos
elbow_flex.pos
wrist_flex.pos
wrist_yaw.pos
wrist_roll.pos
gripper.pos

Training

Base checkpoint:

/root/work/drtc-Phi/outputs/train/rebot_smolvla_flipbread_overfit_20260425_1000steps/checkpoints/001000/pretrained_model

Merged dataset:

phi-media-lab/rebot_flipbreadtopot_20260425_44eps
44 episodes, 18426 frames, 30 FPS

Training command summary:

lerobot-train   --policy.path=outputs/train/rebot_smolvla_flipbread_overfit_20260425_1000steps/checkpoints/001000/pretrained_model   --policy.push_to_hub=false   --dataset.repo_id=phi-media-lab/rebot_flipbreadtopot_20260425_44eps   --dataset.video_backend=pyav   --batch_size=8   --steps=3000   --eval_freq=0   --save_freq=1000   --log_freq=50   --num_workers=4   --wandb.enable=false

Final logged loss: about 0.049.

Validation

The checkpoint reloads with SmolVLAPolicy.from_pretrained(...) and outputs (1, 50, 7) action chunks.

Observed L20 latency:

• first call: about 535 ms
• steady calls: about 151-153 ms/chunk

Safety

This is a reBot-native overfit validation model, not a certified autonomous control policy. Before real actuator execution, use logging-only validation, clipping, rate limits, joint limits, and emergency stop handling.