Walker C1 EDU: Inference Server (Warm Start)¶

Keep the policy resident in a container on the robot side (or GPU side). The operator machine controls it remotely over ZMQ: load a model, hot-swap checkpoints, start/stop execution, return home, and subscribe to status.

Operator inference_client.py ──ZMQ REQ 5690──► inference_server (lerobot venv)
                              ◄──ZMQ PUB 5691──   │
                                                  ├─ lerobot-rollout (26D / 13D)
                                                  └─ ros2_walker_c1_bridge (ZMQ 5661/5662)

1. Scripts¶

File	Purpose
`scripts/deploy/walker_c1/inference_server.sh`	Start `python -m lerobot.scripts.inference_server` in the background (5690/5691)
`scripts/deploy/walker_c1/inference_client.py`	Command client / status subscriber
`scripts/deploy/walker_c1/inference_client.sh`	Thin wrapper (prefers `/lerobot/.venv/bin/python`)

2. Quick start¶

# Robot side (inside the container): confirm/switch to half_sdk (the service validates and refuses otherwise)
bash /ubt_IL/scripts/deploy/walker_c1/sdk_mode.sh half_sdk

# Background + warm start (set WALKER_C1_REAL=1 on hardware to enable the guard)
# Verified on hardware (Nov 2026): 13D real-robot model
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

# Operator side
bash /ubt_IL/scripts/deploy/walker_c1/inference_client.sh status
bash /ubt_IL/scripts/deploy/walker_c1/inference_client.sh start
bash /ubt_IL/scripts/deploy/walker_c1/inference_client.sh home
bash /ubt_IL/scripts/deploy/walker_c1/inference_client.sh stop
bash /ubt_IL/scripts/deploy/walker_c1/inference_client.sh shutdown

launch starts the service and polls until READY:

bash /ubt_IL/scripts/deploy/walker_c1/inference_client.sh launch \
  --policy-path /ubt_IL/model/walker_c1_real_13d_act_1RGB/checkpoints/last/pretrained_model \
  --robot-type walker_c1 --robot-model walker_c1_13d \
  --inference-hz 2 --fps 30

Remote launch (SSH + docker exec on the machine running the container):

bash /ubt_IL/scripts/deploy/walker_c1/inference_client.sh launch --remote \
  --host 192.168.11.3 --ssh-user ubt --ssh-port 22 --container lerobot-tienkung \
  --policy-path /ubt_IL/model/.../pretrained_model \
  --robot-type walker_c1 --robot-model walker_c1_26d

Inference backend recommendation (Walker series): prefer INFERENCE_TYPE=act_async (default). sync is for debugging only and must keep single-action smoothing (WALKER_C1_SINGLE_ACTION_SMOOTH_S, default 0.15 s; setting 0 triggers a warning).

--robot.type is not hard-coded: it uses the robot_type from the load command first, then the ROBOT_TYPE environment variable, and defaults to walker (preserving walker_s2 behavior). The C1 client forwards walker_c1.

3. Client commands¶

Command	Description
`launch`	Start the service (locally or with `--remote`) and wait for READY
`load`	Hot-swap checkpoint / robot model (`--policy-path`, `--robot-type`, `--robot-model`)
`start` / `stop`	Start / stop sending actions (`stop` stops immediately; C1 motion control holds the current pose)
`home`	Smoothly return home (chunk path, 2 s)
`status`	Query status once
`watch`	Subscribe to the PUB status stream (state/loop_hz/chunk_id/robot_connected/uptime_s)
`shutdown`	Shut down the service

4. State machine¶

LOADING ──► READY ⇄ RUNNING
              │         │
              ├──► RETURNING_HOME
              ├──► ERROR
              └──► SHUTTING_DOWN

READY: model loaded, robot connected, no actions being sent
RUNNING: inference and action publishing at inference_hz
RETURNING_HOME: returning home; goes back to READY when finished

5. Environment variables¶

Server (inference_server.sh)

Variable	Default	Description
`POLICY_PATH`	— (required)	`pretrained_model` directory
`CAMERA_TOPIC`	`/sensor/camera/head_stereo_left/color/raw`	Camera topic (C1 real head left; simulation uses `/sensor/camera/head/color/raw`)
`CAMERA_MSG_TYPE`	`shm_msgs/Image6m`	Camera message type (real: `shm_msgs/Image6m`; simulation: `sensor_msgs/Image`)
`ROBOT_TYPE`	`walker_c1`	LeRobot plugin registration name
`ROBOT_MODEL`	`walker_c1_13d`	`ROBOT_MODELS` key (= `--robot.joint_config`)
`ROBOT_CONFIG`	Empty	Explicit hardware config JSON; takes precedence over `ROBOT_MODEL`
`INFERENCE_TYPE`	`act_async`	`act_async` / `sync`
`INFERENCE_HZ`	`1`	Replanning rate (about 1 Hz measured)
`EXECUTION_HORIZON`	`0`	Truncate chunks after N steps (0 = no truncation)
`FPS`	`15`	Control loop FPS
`TASK`	`walker c1 pick place`	Task description
`SERVER_CMD_PORT`	`5690`	REQ command port
`SERVER_STATUS_PORT`	`5691`	PUB status port
`ALLOW_DIM_ONLY_POLICY`	`1`	Precheck dimensions only
`BLEND_HORIZON` / `BODY_PUBLISH_HZ` / `BODY_V_MAX`	Bridge defaults	Chunk blending parameters (passed to the bridge)
`WALKER_C1_SINGLE_ACTION_SMOOTH_S`	`0.15`	`sync` single-action smoothing window (seconds): expands each action into a short current→target trajectory executed at 300 Hz; `0` disables. Recommended 0.15 at FPS=15 and 0.1–0.15 at FPS=30

Client

Variable	Default
`DEPLOY_CONTAINER`	`lerobot-tienkung`
`SERVER_CMD_PORT` / `SERVER_STATUS_PORT`	`5690` / `5691`
`--host`	`127.0.0.1`
`--camera-topic` / `--camera-msg-type`	`/sensor/camera/head_stereo_left/color/raw` / `shm_msgs/Image6m`

launch converts --camera-topic / --camera-msg-type into CAMERA_TOPIC / CAMERA_MSG_TYPE for the server script (for remote launch it prefixes env VAR=... bash ...). The camera is read only at service startup; load does not change the camera source at runtime.

6. FAQ¶

Symptom	Fix
`launch` stays in LOADING	Check `/tmp/walker_c1_inference_server.log`; usually a model path or camera timeout
`status` cannot connect	Check `--host` and ports; is the service in another container?
Old model still used after `load`	Check the `model` field with `watch`; confirm `--robot-model` matches the policy dimension
action dim mismatch	Policy dimension and `ROBOT_MODEL` disagree (26D / 13D); check the training dataset and launch arguments
Robot still moves after `stop`	In-flight chunks stop after at most ~1 tick; if motion continues, confirm the bridge was restarted with the new version (`hold_trajectory` active)
Stale PID file	`/tmp/walker_c1_inference_server.pid`; the singleton guard corrects it automatically