LEMO Platform: From Unboxing to Your First Embodied AI Development

Overview

LEMO is an all-in-one dual-arm development platform for Embodied AI developers, combining teleoperation, data collection, model deployment validation and open development, and covering the full Embodied AI workflow from data collection and model training to inference deployment.

Core hardware: two high-performance A1Z robotic arms, G1Z grippers, a multi-view vision system, and a LEMO-CCU controller.

Product features: ready out of the box, quick assembly and disassembly, and ready to use anywhere.

Target users: tech enthusiasts, universities, research institutes and robotics development teams, for efficient dual-arm manipulation research and Embodied AI application development.

Unboxing & Package Contents

Check all items against the shipping checklist before assembly.

LEMO STUDIO Software

LEMO STUDIO is the supporting software for LEMO. Its main modules are as follows.

Module	Description
Workbench	Device status, current task and collection statistics are viewed in one centralized place.
Task Data	Manage collection progress by task, and link the corresponding Episode data.
Data Collection	The data collection cockpit for teleoperation recording and multi-view camera monitoring.
Model Plaza	Hosts Galaxea self-developed models together with industry open-source models.
Training Task	Create training configurations, view training progress, and enter model deployment.
Model Deployment	Bind the trained model to the physical robot for inference.

LEMO STUDIO is more than a data collection tool—it covers the full pipeline: record demonstrations, prepare task data, select a base model in Model Plaza, launch training in Training Task, and bind the trained model to the robot in Model Deployment. The hands-on example below walks through the first step—recording a cube pick-and-place episode.

Hands-On Data Collection: Pick-and-Place Cube

This example records one complete episode of a cube pick-and-place task.

Step	Action	Operation Notes
1	Start teleoperation	Press the right-hand handle button to start teleoperation; confirm the UI shows "Teleoperating".
2	Start recording	Press the left-hand handle button to start recording; confirm the recording status.
3	Move the follower arm toward the cube	Control the leader arm to drive the follower arm close to the cube.
4	Grasp the cube with the gripper	Press the index-finger trigger to control the gripper and grasp the cube.
5	Move and place the cube	Move to the target position and release the trigger to place the cube.
6	Stop recording	Press the left-hand handle button again to stop recording; wait until saving completes.
7	Reset arms and exit teleoperation	Press the right-hand handle button again to complete reset and end teleoperation.

Training & Inference

LEMO Studio provides the product workflow and compute collaboration solution for robot skill training and inference: collect and filter task data, then train and deploy models. Available models include Galaxea self-developed models (G series, π series, Lingbot-VLA) and industry open-source models.

Platform Architecture & Workflow

Training and inference can use different compute resources; the robot and the service are associated through independent sessions.

1. Task data: collect demonstrations, filter and convert.

2. Model training: base model and dataset, LoRA or full fine-tuning.

3. Inference deployment: connect to the model service, two-way bind the robot.

4. Robot execution: upload task observations, receive actions.

Robot side: collect images and state, receive and validate actions; two-way confirm identity and session with the model service.

Compute side: edge compute, cloud compute and Galaxea cloud compute, responsible for model training and online inference respectively.

Task Data & Training Dataset

Trace the collection source by task; the training entry selects the converted LeRobot dataset.

1. Task Filtering: view Episodes by task.

   

2. Quality Check: view completeness and the playback entry.

   

3. LeRobot Conversion: select data and configure the task.

   

Model Training & Open-Source Models

LoRA updates the adaptation parameters; full fine-tuning updates the complete model parameters. The exact supported scope depends on model adaptation. Configure dataset, compute and training method, with advanced parameters and recommended defaults. Available entries: Galaxea G series, π0 / π0-Fast / π0.5, Lingbot-VLA, and custom model entry.

Three Types of Compute

The three compute types are matched on demand; the actual training method depends on model adaptation and resource configuration.

Model Deployment & Two-Way Robot Binding

Two-way confirm device identity, model version and session information; isolate observations and actions by session. Connect to the model service, select the target robot, and execute after the two-way handshake. The device number is read automatically and session information is generated automatically.

Training-Inference Delivery Pipeline

Each stage retains traceable deliverables; final acceptance is based on real-robot task performance.

If model performance or task results do not meet the target, return to data filtering and training configuration to continue iterating.

Takeover & Real-Robot Reinforcement Learning

When the model makes a mistake, a human can take back control with one key. This "human correction" data is the most valuable part of real-robot reinforcement learning. LEMO is a dual-arm tabletop manipulation platform for Human-in-the-Loop RL.

Why Human Takeover Is Needed

A model deployed on a real robot will inevitably encounter unseen situations; fallback and correction both rely on humans.

• The model has boundaries: out-of-distribution scenes such as lighting, occlusion, or object positions beyond the training distribution may cause drift, jitter, or even collision against the limit. A mechanism to take back control at any time is needed.

• The human is the best corrector: one-key takeover (model stops sending, human drives directly); jump-free alignment (smoothly align to the hand pose before handing over); the correction action itself is a high-quality "correct answer".

• Correction data is the most valuable: the takeover segment is a paired sample of "what the model got wrong at the time" and "what the human did right", covering an order of magnitude more efficiently than recollecting teleoperation data; it is the core fuel for online/offline reinforcement learning.

How Takeover Works

From the operator's view: one key press, four steps.

1. Initiate: while inference (POLICY) is running, press the takeover key on the handle or the "Takeover" button on the page.

2. Align: the model stops sending; the teleoperation arm moves to the current isomorphic pose of the robot arm.

3. Take over: the human directly drives the dual arms; model output is shown read-only and no longer sends control.

4. Return / end: press again to hand control back to the model for continued inference, or end this session.

• Takeover can only be initiated during inference. Pressing it in other states (teleoperation / standby / homing) has no effect, to prevent accidental triggers; exiting takeover is always allowed, as finishing up must always work.

• Automatic lock on abnormality: link packet loss, arbitration loss, or heartbeat timeout automatically enters PAUSE and holds the pose (fail-closed).

How Takeover Data Is Recorded and Used in Training

Inference/takeover segments are recorded automatically and physically isolated from pure teleoperation data.

1. Inference running: POLICY mode, the model drives the dual arms.

2. Auto recording: recording starts upon entering POLICY; feedback, execution target and model original image are labeled frame by frame.

3. Takeover correction: the TAKEOVER segment continues recording in the same episode, labeling control_src frame by frame (model segment / human segment separable).

4. Isolated storage: independent dataset directory; each episode carries the takeover ratio, with zero confusion with teleoperation data.

Each data frame records two actions simultaneously: the actually executed action (what the arm really did) and the model's intended action (raw output) — the divergence curve is clear at a glance.

Training-side usage: filter takeover segments by control_src for correction learning; assess model weaknesses by takeover ratio; freely combine pure teleoperation, pure inference and mixed data.

Real-Robot Reinforcement Learning Loop

Target form: deploy → takeover correction → data return → online update → redeploy.

1. Deploy & inference: the model works on the real robot; the dual arms operate according to model commands.

2. Takeover correction: a human takes over and corrects at any time; what the model got wrong and what the human did right are both precisely recorded.

3. Data return: correction segments are uploaded to the training server, each recorded exactly once, without duplication or omission.

4. Online update: the server updates the model with new data, using the time when the robotic arm returns to zero, without occupying the working rhythm.

The loop returns to "deploy" — the new model is sent down again. The model becomes smarter the more it runs on the real robot, and the number of human takeovers needed keeps decreasing.

Ready to Start Building?

LEMO brings dual-arm embodied AI development out of the lab and onto your desk—unbox it, collect your first demonstration, train your first policy, and watch it run on real hardware, all in one integrated platform.

Whether you're a research team, a robotics lab, or a product team exploring embodied AI, LEMO gives you the full stack—hardware, teleoperation, data collection, model training, and real-robot deployment—ready out of the box.