What Is Robot Data Collection?

Robot data collection (embodied-AI data collection) is the process of synchronously recording paired data — observations (camera images, joint states) plus actions (joint commands) — while a robot performs a task, typically demonstrated by a human via teleoperation. This data is the raw material for training VLA models and imitation-learning policies: what a model can learn is bounded by what was collected.

The Leader-Follower Teleoperation Pipeline

Leader-follower arms are a common real-robot collection setup:

1. The operator moves a lightweight leader arm; the follower arm mirrors its joint motion in real time;
2. The system records multi-camera images, follower joint states, and action commands at a fixed rate (typically 30-50 Hz or higher);
3. Each completed task is saved as one trajectory (episode); a task usually needs tens to hundreds of episodes;
4. Data is filtered and aligned, then fed into the training pipeline.

When leader and follower correspond joint-to-joint, demonstration is intuitive and can reduce online kinematic-mapping complexity. The BXI UpperBody 1 dual-arm teleoperation platform uses a dual-leader, dual-follower architecture and can integrate with ROS2 recording workflows; available interfaces and data formats depend on the project configuration.

Real-Robot Data vs Simulation Data

• Real-robot data captures true friction, deformation, lighting, and sensor noise, so trained policies deploy directly — but collection is labor-intensive and costly;
• Simulation data scales cheaply and in parallel, but differs from reality and must be bridged via Sim-to-Real techniques.

In practice the two are mixed: simulation provides scale, real data anchors the true distribution.

How Imitation Learning Consumes the Data

The ALOHA/ACT line of work trains policies directly on leader-follower bimanual trajectories: ACT (Action Chunking with Transformers) predicts a short chunk of future actions as a unit. The required demonstration count depends on task difficulty, data quality, and the desired generalization range. These methods make teleoperated data collection plus imitation learning a practical route for training embodied-AI policies.