VLAI L1
L1 高性能移动操作双臂机器人

What is the VLAI L1?

The VLAI L1 (L1 高性能移动操作双臂机器人) is a mobile dual-arm manipulation platform designed by 深圳未来动力有限公司 (VLAI) and manufactured in China. RoboticsCenter distributes and supports the L1 in the United States, providing warranty service, SDK access, and integration with the Fearless Data Platform.

The L1 combines a two-wheel differential mobile base with a vertical linear lift and two 7-DOF arms — each equipped with a self-developed 8Nm gripper — giving researchers and developers a compact, high-accuracy platform for imitation learning, VLA training, and production teleoperation workflows.

Key differentiators

• ±0.02mm repeat positioning accuracy via dual-encoder feedback and MIT motor protocol with FOC control.
• <10ms end-to-end control latency over CAN-FD at 5Mbps — low enough for real-time force-sensitive manipulation.
• 16 DOF total: 7 DOF + 1 gripper per arm, 6kg payload per arm (12kg combined), 63cm arm span.
• Adjustable height from 106cm to 162cm via motorized linear lift at 30mm/s — covers table, shelf, and floor-level tasks without repositioning.
• Simulation-first design: Isaac Sim, Isaac Lab, and MuJoCo environments ship with the Developer tier and above, enabling rapid sim-to-real transfer.
• Native ROS2 + MoveIt2 — no middleware translation layer needed.

Hardware at a glance

Get running in three steps

1. Purchase your tier Order the L1 through RoboticsCenter at roboticscenter.ai/contact. We recommend the Developer tier for full SDK access, or Developer Pro if you need VR teleoperation from day one. Lead time is typically 6–8 weeks for US delivery.
2. Install the SDK and connect Install the roboticscenter Python package and connect to the robot over your local network. The CLI opens a browser session automatically.
   bash Copy

   pip install roboticscenter
   rc connect --device l1
   # Terminal: Session ready → https://platform.roboticscenter.ai/session/RC-XXXX-XXXX

3. Open the browser teleop panel Navigate to the session URL printed in your terminal. The panel shows both arms in real time, base WASD controls, lift sliders, and a live frame counter. All data is streamed and auto-uploaded to your platform workspace.

Choosing your L1 tier

The L1 ships in four tiers. All share the same mechanical frame and gripper hardware; the differentiators are the onboard compute controller, bus protocol, bundled software, and warranty.

★ Developer Pro is the recommended entry point for most research and small-team production deployments. Prices are approximate USD at current exchange rates; final quotes provided by RoboticsCenter.

Installation & connection

Requirements

• Python 3.10+ (3.11 recommended)
• L1 powered on and connected to the same local network as your workstation (Ethernet or 5GHz Wi-Fi)
• Developer tier or above (Youth tier does not include SDK access)

Install

pip install roboticscenter

Connect and stream frames

from roboticscenter import L1Robot

robot = L1Robot.connect()
print(robot.session_url)

for frame in robot.stream():
    joints = frame.data['joints']
    print(f"Left: {joints['left_arm']}")
    print(f"Right: {joints['right_arm']}")
    print(f"Base: {frame.data['base']}")

Mock mode (no hardware required)

Mock mode replays a synthetic joint trajectory and camera feed. The session URL, frame format, and all SDK methods are identical to live hardware — making it safe to develop and test data pipelines before your robot ships.

# Test without hardware
rc connect --device l1 --mock

ROS2 + MoveIt2

The L1 runs a ROS2 node on its onboard computer. On Developer tier and above, you can subscribe to all standard joint and sensor topics from your workstation on the same network, or bridge them to the cloud via the robot agent.

Check available topics

ros2 topic list

Key ROS2 topics

Subscribe to joint states

ros2 topic echo /joint_states

Launch the robot agent (cloud bridge)

Run the agent on the L1's onboard computer to bridge ROS2 topics to the platform session, enabling remote teleoperation and cloud data collection.

# ROS2 agent (on robot's onboard computer)
python l1_robot_agent.py \
  --backend wss://platform.roboticscenter.ai \
  --session RC-XXXX-XXXX \
  --ros2

Browser teleop panel

Open the session URL from rc connect to access the full browser-based teleop panel. No additional software installation is needed on the operator's machine.

Panel overview

• Dual arm view — Real-time 3D visualization of both 7-DOF arms. Joint angles, end-effector pose (6D), and gripper state are displayed at <10ms update rate from the CAN-FD bus.
• Base controls (WASD) — Keyboard or on-screen WASD drives the differential base at up to 2m/s. Speed scaling is adjustable via a slider (10%, 50%, 100%).
• Lift controls — Vertical slider adjusts lift height between 106cm and 162cm in real time at 30mm/s.
• End-effector mode selector — Switch the active end-effector type per arm independently: Gripper (default 8Nm), Dexterous hand, or Suction cup. Switching updates force/position limits accordingly.
• Camera feeds — Chest camera stream (Developer+), and left/right wrist streams (Developer Max). Streams are displayed in-panel with configurable resolution.
• Frame counter — Live frames-per-second and total frames captured in the current session, updated every 500ms.

VR teleop (Developer Pro and Max)

Developer Pro and Max tiers unlock VR headset control. Plug in an OpenXR-compatible headset and run rc vr --session RC-XXXX-XXXX to map hand controller pose directly to each arm's end-effector in Cartesian space, with haptic feedback proportional to estimated contact force.

Session-based data collection

Every rc connect session is a named data collection unit. Sessions auto-upload to your Fearless Platform workspace on close — no manual export step required.

What gets captured per frame

• Joint data — Position, velocity, and effort for all 16 DOF at the CAN-FD polling rate (~500Hz). Stored as float32 arrays with microsecond timestamps.
• Camera streams — Chest RGB (all tiers with camera), wrist RGB (Developer Max). Each frame is timestamped and synchronized to joint data within ±1ms.
• Force / contact signals — Gripper jaw force estimate from motor current, updated at joint polling rate.
• Base state — Wheel odometry, linear and angular velocity, lift height.
• Operator actions — Raw teleop command stream (keyboard, VR controller, or programmatic), stored alongside robot state for imitation learning.
• Episode metadata — Session ID, tier, firmware version, task label (if set), start/end timestamps, and frame count.

Labeling during collection

Press Space in the browser panel to mark episode boundaries. Press L to annotate the current frame with a custom label string — useful for tagging successful grasps, failure events, or task transitions in real time.

From session to trained policy

After a collection session ends, the Fearless Platform provides a one-click pipeline that takes raw session data through cleaning, annotation, and policy training — all without leaving the browser.

1. Clean Automatic filtering removes frames with sensor dropout, joint limit violations, or camera sync errors. A summary report shows how many frames were retained.
2. Annotate Apply task labels, success/failure flags, and episode boundary corrections. The annotation studio shows synchronized joint, camera, and force timelines side-by-side.
3. Train VLA / policy model Select a base model (ACT, Diffusion Policy, or custom VLA checkpoint) and launch a training run on the platform's GPU cluster. Training progress and eval metrics are streamed back to the browser.

Dataset export

Cleaned and annotated sessions can be exported in LeRobot format (HDF5 + JSON manifest), RLDS, or raw JSONL + MP4. Use the Data tab in the platform to configure export format and download or sync to external storage.

Communication architecture

CAN-FD bus (onboard)

The L1 uses CAN-FD at 5Mbps (Developer, Pro, Max) between the onboard controller and all joints. CAN-FD's larger payload frame (up to 64 bytes vs. CAN2.0's 8 bytes) allows batching position, velocity, and torque data for all joints in fewer bus cycles, achieving the <10ms control loop latency. The Youth tier runs CAN2.0 which is sufficient for lower payload (2kg) operation but not recommended for high-frequency force-sensitive tasks.

ROS2 topics (LAN)

The onboard ROS2 node publishes the full topic list described in the ROS2 section above. Topics are available on the local network without any additional configuration once the robot is powered on.

WebSocket session relay (cloud)

The robot agent connects to the platform via WebSocket (wss://platform.roboticscenter.ai) and relays the following per session:

• Bidirectional control commands (teleop → robot)
• Outbound sensor streams (robot → platform) with per-frame sequence numbers for loss detection
• Session control signals: start, pause, episode-boundary, label, stop
• Heartbeat at 1Hz; reconnect with exponential back-off on disconnect

Common errors & fixes

What teams build with the L1

The L1's combination of mobile base, adjustable lift, dual 6kg-payload arms, and native simulation tooling makes it well-suited for a wide range of manipulation-centric deployments.