Stage 6: Model Deployment (Linux)

This stage loads the trained policy so the robot can execute tasks autonomously, and records evaluation videos to verify the results. This is the finale of the whole workflow and the key test of the training outcome.

---

Prerequisites

• Stage 5: Model Training completed

• Training produced outputs/train/soarm_amazing_hand_pick/checkpoints/last/pretrained_model/

• The camera indices are recorded

---

Step 1: Confirm the Model Files

ls outputs/train/soarm_amazing_hand_pick/checkpoints/last/pretrained_model

It should contain model files such as model.safetensors.

⚠️ Note (model path): --policy.path must point to the pretrained_model directory (containing the config + weights), not the checkpoint root directory.

---

Step 2: Deploy and Evaluate

lerobot-rollout \
  --strategy.type=episodic \
  --robot.type=so101_amazing_hand \
  --robot.port=<follower_arm_port> \
  --robot.hand_port=<hand_port> \
  --robot.id=amazing_hand_follower \
  --robot.cameras='{
    wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30, fourcc: "MJPG"},
    top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30, fourcc: "MJPG"}
  }' \
  --policy.path=outputs/train/soarm_amazing_hand_pick/checkpoints/last/pretrained_model \
  --dataset.repo_id=rollout_soarm_amazing_hand_pick_eval \
  --dataset.root=~/lerobot_data \
  --dataset.push_to_hub=false \
  --dataset.num_episodes=10 \
  --dataset.single_task="Pick up the cube with the dexterous hand" \
  --display_data=true

Replace <follower_arm_port> / <hand_port> with the actual paths; replace the camera index_or_path with your camera indices.

💡 Notes: Use lerobot-rollout but do not add ****--teleop.type; the policy then controls the robot autonomously (replacing manual teleoperation). The data is saved as an evaluation set. --dataset.root / --dataset.push_to_hub=false are the same as in Stage 4; purely local saving requires no HF login.

---

Evaluation Procedure

1. Return the robot + hand to the starting position

2. Press Enter to start: the policy executes the task autonomously

3. Observe whether the grasp succeeds (press Enter to continue after each episode)

4. Repeat for num_episodes episodes

Evaluation metric: success rate = successful episodes / total episodes

⚠️ Note 1 (reset consistency): Start every episode from the same starting position, otherwise the policy fails to generalize and the success rate will be artificially low.

⚠️ Note 2 (safety): On the first autonomous run, it is recommended to keep a hand on the robot / run slowly and observe, to confirm the policy's motions are reasonable. The policy may make unexpected movements.

⚠️ Note 3 (expected success rate): ACT typically achieves a 50-80% success rate with 20 episodes of data. If it is lower than expected, go back and record more data or adjust the training step count.

⚠️ Note 4 (headless environment): --display_data=true requires a display server; in a GUI-less environment, remove this parameter (evaluation still runs, it just won't display in real time).

---

Iterative Optimization

If the evaluation success rate is unsatisfactory, adjust in order of priority:

Priority	Optimization	Action
1	Record more high-quality data	Go back to Stage 4 and record an additional 20-30 more consistent episodes
2	Increase the training step count	Go back to Stage 5, --steps=100000
3	Check starting-position consistency	Strictly reset before every evaluation episode
4	Adjust the task description	Make sure single_task matches the task

---

This completes the full closed loop of SO-ARM101 + AmazingHand: calibration → teleoperation → collection → training → deployment.

---

Troubleshooting

Symptom	Cause	Solution
Model fails to load	Wrong/incomplete path	Confirm --policy.path points to the pretrained_model directory
Policy does not move	Camera/observation error	Confirm the camera indices match those at training time; check /dev/video* permissions
Policy moves erratically	Inconsistent starting position / poor data	Reset strictly; record more data
Behavior differs from training	Environment differences	Confirm the cameras, lighting, and object positions match those at recording time

Related Products

SO-ARM101 Developer Kit→AmazingHand 4-Finger Dexterous Hand→