Deploy¶

Robot: Walker C1 EDU Data source: real-robot teleoperation/scripted collection (Thinker Studio or ubt_sim/teleoperation/control/walker_c1) Code: ubt_IL/scripts/{convert,train,eval,deploy}/walker_c1/

Simulation and the real robot share the same ROS 2 interfaces (mc_task_msgs / mc_state_msgs / sensor_msgs/Image) and are isolated by ROS_DOMAIN_ID. The policy deployment path sends only 26 dimensions (both arms + both hands) or 13 dimensions (right arm + right hand); the head/waist/legs hold their current positions.

✅ Verified on hardware (Nov 2026): the full loop works with 13D = right arm 7 + right hand 6 (ROBOT_MODEL=walker_c1_13d, 30 fps, head camera head_stereo_left/color/raw 640×512): dataset walker_c1_real_13d_1RGB (34 episodes / 11,406 frames) -> model walker_c1_real_13d_act_1RGB -> real-robot deployment.

Real robot ROS 2 ──► ubt_sim/teleoperation/control/walker_c1 (collection) ──► HDF5
                                                                             │
Container: convert.sh → train.sh → eval_policy.py → robot_ready.sh + rollout.sh
                                                                             │
LeRobot (3.12) ──ZMQ 5661/5662──► ros2_walker_c1_bridge.py (3.10) ──► real robot

0. Workflow overview¶

Stage	Location	Command
Switch SDK control	Operator machine / container	`bash scripts/deploy/walker_c1/sdk_mode.sh half_sdk`
Real-robot collection	Operator machine / ubt_sim container	`teleoperation/control/walker_c1/save_data.sh`
Container	Operator machine	`cd ubt_IL/docker && bash run.sh start && bash run.sh bash`
Conversion	Container	`scripts/convert/walker_c1/convert.sh`
Training	Container (GPU)	`scripts/train/walker_c1/train.sh`
Evaluation	Container	`scripts/eval/eval_policy.py`
Deployment	Container (robot side)	`robot_ready.sh` + `rollout.sh`

1. Real-robot prerequisites¶

Switch to SDK control (half_sdk required): C1 arms accept ROS 2 joint commands only in workMode=develop + developModeState=HalfSdkActive (normal ignores commands; full_sdk hands over whole-body balance and is not used for VLA). The deployment script performs the full sequence (enter developer mode → legs STAND → half_sdk) over SSH in the robot's developer container (ubt@192.168.11.3:2222):

bash /ubt_IL/scripts/deploy/walker_c1/sdk_mode.sh status     # work mode / motion permission / legs / e-stop
bash /ubt_IL/scripts/deploy/walker_c1/sdk_mode.sh check      # validate only (0 = deployable)
bash /ubt_IL/scripts/deploy/walker_c1/sdk_mode.sh half_sdk   # switch
bash /ubt_IL/scripts/deploy/walker_c1/sdk_mode.sh normal     # restore normal

rollout.sh / robot_ready.sh / robot_home.sh / inference_server.sh include the guard and refuse to run outside half_sdk (on hardware set WALKER_C1_REAL=1 or SDK_GUARD=1; simulation uses the default SDK_GUARD=auto and skips the check).

ROS domain: the real robot defaults to ROS_DOMAIN_ID=0; simulation and real deployment must use the same value. Set both to 146 when isolation is needed.
DDS: Thor (C1 onboard compute) must use the same RMW as the main control stack — the robot stack uses rmw_cyclonedds_cpp + CYCLONEDDS_URI (bound to mgbe2_0, local 192.168.11.3). docker/run.sh injects the same environment and mounts docker/cyclonedds_thor.xml (CYCLONEDDS_URI=file:///opt/cyclonedds_thor.xml), so no manual setup is needed. Two field-tested caveats when editing the config:
<Interfaces> must be inside <General> (Humble's CycloneDDS 0.10+). Under <Domain> it reports unknown element and falls back to the default NIC. List only mgbe2_0, not lo (not multicast).
Domain 0 has many participants; the default limit raises Failed to find a free participant index for domain 0. The config sets ParticipantIndex=none and MaxAutoParticipantIndex=120.

For FastDDS (simulation/remote machines), set RMW_IMPLEMENTATION=rmw_fastrtps_cpp and add the local NIC IP (Thor 192.168.11.3, Walker peer 192.168.11.99) to the docker/fastdds_no_shm.xml whitelist.

State self-check (inside the container, after run.sh bash):

source /opt/ros/humble/setup.bash
source /ubt_IL/walker_s2/walker_sdk_ros2/install/setup.bash
export ROS_DOMAIN_ID=0
ros2 topic hz /mc/sdk/robot_state                                 # measured ≈ 460-490 Hz on Thor
ros2 topic hz /sensor/camera/head_stereo_left/color/raw           # shm_msgs/Image6m
ros2 topic info -v /mc/sdk/robot_command                          # should have subscribers (SDK control active)

The real-robot RobotState contains only 19 joints (14 arm + 2 head + 3 waist, no legs). The bridge selects the 14 arm joints by name, so it is unaffected; head/waist/legs are not part of the policy and are not commanded.

Safety: place the robot in a safe position; use DURATION=10 for the first short validation; keep the emergency stop within reach.

2. Real-robot collection¶

# Inside the ubt_sim container (same domain as the robot; do not reuse simulation processes)
bash teleoperation/control/walker_c1/save_data.sh
bash teleoperation/control/walker_c1/save_data.sh --batch 20

The output matches simulation: ubt_sim/dataset/walker_c1/<epoch>/trajectory.hdf5 (15 Hz, 31 body joints + 6+6 hand channels, one RGB stream). You can also use the Thinker Studio teleoperation platform and export a LeRobot v3 dataset directly.

The collection side needs the real camera topics — C1 hardware does not publish /sensor/camera/head/color/raw; it provides:

Topic	Type	Notes
`/sensor/camera/head_stereo_left/color/raw`	`shm_msgs/Image6m`	Policy input: 1920×1536, yuv422 (UYVY), step=3840, fixed 6 MB
~~`/sensor/camera/head_stereo_left/color/compressed`~~	`sensor_msgs/CompressedImage`	No longer published after the onboard vision stack restart
`head_stereo_right` / `head_fisheye_left,right` / `mono_chin` / `wrist_left,right`	Same family	Other views

Aspect ratio: native 1920×1536 (5:4). The collection side stores JPEG at half size (960×768, still 5:4). Training and deployment consistently use 640×512 (5:4, exactly 1/3 of 1920×1536) with no cropping and no aspect-ratio change.

Use WALKER_C1_CAMERA_TOPIC + WALKER_C1_CAMERA_MSG_TYPE for collection/deployment (the deployment script uses CAMERA_TOPIC / CAMERA_MSG_TYPE). Topics ending in /compressed are inferred as compressed images; everything else defaults to sensor_msgs/Image. The real shm_msgs/Image6m type must be specified explicitly, otherwise the subscription type does not match.

3. Start the container¶

cd ubt_IL/docker
bash run.sh build && bash run.sh start && bash run.sh check && bash run.sh bash

Thor (L4T r38 / sm_110) is selected automatically by env.sh, which uses Dockerfile.thor (image lerobot-tienkung:humble-thor, based on the robot's ros2_ws:humble-dyn). Do not use the Orin Dockerfile.arm64 (r36 base + sm_87 wheels; it does not run on Thor).

run.sh check should include:

[OK] walker_c1 plugin: installed
[OK] walker_c1 bridge: available
[OK] walker ROS2 msgs: installed (mc_state_msgs / mc_task_msgs, sharing the workspace with walker_s2)
[OK] torch: 2.9.0+cu130 | arch: ...,sm_110,... | kernel: ok
[INFO] RMW: rmw_cyclonedds_cpp | CYCLONEDDS_URI=file:///opt/cyclonedds_thor.xml

4. Data conversion (HDF5 -> LeRobot)¶

4.1 Conversion command¶

Verified path: real-robot 13D (config configs/walker_c1_real_13d_1RGB.json, 13D = right arm 7 + right hand 6, head camera 640×512, 30 fps):

CONFIG=/ubt_IL/scripts/convert/walker_c1/configs/walker_c1_real_13d_1RGB.json \
SRC_ROOT=/ubt_IL/dataset/walker_c1 \
REPO_ID=walker_c1_real_13d_1RGB \
FPS=30 \
bash /ubt_IL/scripts/convert/walker_c1/convert.sh --overwrite

If the data is still in ubt_sim/dataset/walker_c1, omit SRC_ROOT (the script falls back automatically) or set it explicitly.
Real-robot recordings are 30 fps; keep the conversion FPS equal to the dataset frame rate. Use RESAMPLE_FPS for other rates.
For 26D real-robot data use walker_c1_real_26d_1RGB.json (dual-arm task).
Real hand ranges were verified: right hand thumb_swing ≤0.68 / thumb_mcp ≤0.82 / four fingers ≤0.85, all within deployment limits (0.96 / 1.04 / 1.57).

For simulation conversion see Simulation workflow · Data conversion.

4.2 Data visualization¶

HF_HUB_OFFLINE=1 lerobot-dataset-viz \
  --repo-id walker_c1_real_13d_1RGB \
  --episode-index 0 \
  --root /ubt_IL/dataset/walker_c1_real_13d_1RGB

Walker C1 EDU dataset preview

Check especially: image resolution (3×512×640), state/action dimension and order (13D = right arm 7 + right hand 6), hand range, and empty episodes.

5. Training and evaluation¶

Verified path: real-robot 13D training (config train_config_walker_c1_real_act_13d_1RGB.json, dataset walker_c1_real_13d_1RGB, output walker_c1_real_13d_act_1RGB):

CONFIG=/ubt_IL/scripts/train/walker_c1/configs/train_config_walker_c1_real_act_13d_1RGB.json \
OUTPUT_DIR=/ubt_IL/model/walker_c1_real_13d_act_1RGB \
bash /ubt_IL/scripts/train/walker_c1/train.sh

train.sh defaults to the simulation 26D config; real-robot training must set CONFIG= explicitly. When real lighting differs strongly from simulation, image_transforms is enabled by default in the config (color jitter / random affine).

Offline evaluation (real-robot dataset):

/lerobot/.venv/bin/python /ubt_IL/scripts/eval/eval_policy.py \
  --policy-path /ubt_IL/model/walker_c1_real_13d_act_1RGB/checkpoints/last/pretrained_model \
  --dataset-path /ubt_IL/dataset/walker_c1_real_13d_1RGB \
  --episodes 5 --inference-freq 1 --plot \
  --plot-dir /ubt_IL/scripts/eval/output/eval_walker_c1_real_13d_act \
  --output /ubt_IL/scripts/eval/output/eval_walker_c1_real_13d_act/results.json \
  --device cuda

With small real datasets, inspect per-dimension curves for plausibility (hand jitter, wrist overshoot) instead of relying only on the total MSE.

6. Deployment¶

Real-robot deployment (on-site photo):

Walker C1 real-robot deployment

# ⓪ Switch to half_sdk (developer mode + half-body SDK; idempotent, only prints state if already ready)
bash /ubt_IL/scripts/deploy/walker_c1/sdk_mode.sh half_sdk
bash /ubt_IL/scripts/deploy/walker_c1/sdk_mode.sh check      # 0 = deployable

# ① Camera check (real: head_stereo_left/color/raw, shm_msgs/Image6m)
python3 /ubt_IL/scripts/deploy/walker_c1/preview_camera.py --once --save-frame /tmp/c1.jpg \
  --msg-type image6m --topic /sensor/camera/head_stereo_left/color/raw

# ② Move to the ready pose (dry-run first to inspect targets and stage durations)
bash /ubt_IL/scripts/deploy/walker_c1/robot_ready.sh --dry-run
bash /ubt_IL/scripts/deploy/walker_c1/robot_ready.sh

# ③ Single-episode rollout (WALKER_C1_REAL=1 enables the half_sdk guard; refuses outside half_sdk)
#    Verified path: real-robot 13D model (on hardware ROBOT_MODEL defaults to walker_c1_13d)
WALKER_C1_REAL=1 SDK_GUARD=1 \
POLICY_PATH=/ubt_IL/model/walker_c1_real_13d_act_1RGB/checkpoints/last/pretrained_model \
DURATION=10 FPS=30 RESET_MODE=task \
CAMERA_TOPIC=/sensor/camera/head_stereo_left/color/raw \
CAMERA_MSG_TYPE=shm_msgs/Image6m \
  bash /ubt_IL/scripts/deploy/walker_c1/rollout.sh

# ④ Return home (home = all 14 arm joints at 0, arms straight down, hands open)
bash /ubt_IL/scripts/deploy/walker_c1/robot_home.sh --dry-run
bash /ubt_IL/scripts/deploy/walker_c1/robot_home.sh

# ⑤ Optional warm-started inference server (see inference-server.md)
WALKER_C1_REAL=1 \
POLICY_PATH=/ubt_IL/model/walker_c1_real_13d_act_1RGB/checkpoints/last/pretrained_model \
  bash /ubt_IL/scripts/deploy/walker_c1/inference_server.sh

Ready pose includes a head-down angle of head_pitch=0.65 rad (37.2°): field tests show 0.50 only sees the table edge, while 0.65 covers the full table. Adjust with --head-pitch; --no-head leaves the head untouched. The three stages (raise shoulder → fold elbow → ready pose) default to ≥ 5 s each (15 s total + 0.5 s hold), and the log prints peak displacement and velocity (≈0.5 rad/s).
home is the exact reverse of the three ready stages (wrist/shoulder pitch to zero → unfold elbow → lower shoulder); measured final error max |pos| ≈ 0.02 rad. It moves only the arm and hand; head/waist/legs are untouched.
Inference backend: prefer act_async (default); serial sync is for debugging only and must keep single-action smoothing (default 0.15 s; setting 0 triggers a warning).
RESET_MODE: the real robot has no /sim/cmd_reset; use task (pose only) or none.
max_relative_target: null in the config JSON; set e.g. 0.05 rad/step when per-step limiting is needed.
disable_torque_on_disconnect=true: smoothly return home on exit (2 s chunk).

Real-robot deployment demo:

See Real-robot deployment for the full guard, registry, and safety precheck details.

7. FAQ¶

Symptom	Fix
Arm does not move at all	Run `sdk_mode.sh check`: is it `develop + HalfSdkActive + standing`? Then check whether `/mc/sdk/robot_command` has subscribers and whether `ROS_DOMAIN_ID` matches
Only the right hand moves	The left hand is all zeros in the data (task characteristic); the policy reproduces this, which is normal
Joint 4 direction/name mismatch	SDK docs call it `elbow_roll`, the URDF calls it `elbow_pitch`; the bridge tolerates aliases, but verify the sign on hardware
Sudden motion	Set `max_relative_target`; lower `FPS`; check hand limits
Camera artifacts / no image	Simulation `sensor_msgs/Image` encoding should be `rgb8`; real is `shm_msgs/Image6m` (UYVY); QoS must be BEST_EFFORT
Arm does not return after exit	`disable_torque_on_disconnect` was set false, or the bridge was killed externally
Inference service unreachable	Ports 5690/5691; check the state machine with `inference_client.sh status`
Empty topic list / no `/mc/*`	RMW mismatch: on hardware use `rmw_cyclonedds_cpp` + `cyclonedds_thor.xml` (see §1 item 3)
`Failed to find a free participant index`	CycloneDDS default participant limit is too low; use `ParticipantIndex=none` + `MaxAutoParticipantIndex=120`
Camera topic missing / no frames	On hardware use `CAMERA_TOPIC=.../head_stereo_left/color/raw` + `CAMERA_MSG_TYPE=shm_msgs/Image6m`; green/purple colors mean UYVY is being decoded as YUYV
`torch.cuda.is_available()` is False	The container lacks `NVIDIA_VISIBLE_DEVICES=all` / `NVIDIA_DRIVER_CAPABILITIES=all` (`run.sh start` sets them)