Dual xArm robot arms coordinate to manipulate a bag during a bimanual imitation learning task

UFACTORY Where Robot Imitation Learning Starts

Home Applications Imitation Learning

Background

The rise of embodied AI is reshaping how robots are developed. Instead of writing explicit code for every task, researchers and engineers now teach robots by demonstration — a paradigm known as Imitation Learning (IL). At the heart of this approach is data: the quality and diversity of demonstration data directly determines the upper bound of a robot’s capabilities.

xArm 7 with a camera-mounted reacher-grabber moves a cup to the red target area during a tabletop manipulation task

Industry Pain Points

The Hardware Gap in
Early-Stage Development

Dual-arm systems and humanoid robots carry enormous hardware costs and complex calibration requirements that make them impractical for early R&D. The gap UFACTORY fills: Performance capable enough to support serious IL research. Price point low enough to deploy at scale. UFACTORY cobots are the practical choice for this space.

Why UFACTORY Cobots

 

Low-Cost Data
Collection Array

At a fraction of the cost of comparable arms, UFACTORY lets you deploy a parallel data collection array — more robots, more data, same budget.

Seamless Integration
with the IL Stack

Connects via Python SDK to work directly with LeRobot — supporting ACT, Diffusion Policy, and Flow Matching out of the box. ROS 2 is also supported for teams that need it.

Low-Latency Joint
Execution

UFACTORY's control loop delivers consistently low command latency, ensuring that action sequences from trained policies translate into smooth, stable motion on the physical arm.

Imitation learning with
multimodal data

We’ve run the full IL pipeline — teleoperation, data collection, training, and inference — on UFACTORY xArm 7 using both Gello and Pika teleoperation hardware with the LeRobot framework. All code and collection guides are open-sourced on GitHub.

Validated IL Pipeline

UFACTORY cobots has been validated end-to-end with the most widely used open-source IL tools. All datasets and collection guides are open-sourced on GitHub.

Dual xArm 6 + UMI + LeRobot ACT

A dual xArm 6 setup uses UMI teleoperation to collect bimanual demonstrations and train a LeRobot ACT policy for coordinated manipulation. The dataset includes 50 UMI teleoperation episodes, with the ACT policy trained for 200,000 steps.

xArm 7 + Gello + LeRobot ACT

Gello's joint-by-joint mapping to UFACTORY xArm 7 captures operator intent with high accuracy, producing clean trajectory data well-suited for precise manipulation tasks. 80 episodes, trained with ACT over 500,000 steps

xArm 6 + Pika + LeRobot ACT

Pika works with any robot arm out of the box, making it a faster option for teams that frequently switch between environments. The dataset includes 150 episodes, with the ACT policy trained for 400,000 steps.

 

Two xArm robotic arms coordinate to manipulate a bag during a bimanual imitation-learning task



An xArm holds a bunch of grapes above a cardboard box during a teleoperated imitation-learning demonstration



An operator guides an xArm using a handheld teleoperation controller



Related Case Studies

Imitation Learning Research

UC Berkeley

Researchers collected demonstration data using a commercially available assistive grabber. The trained policy achieved 87% success on pushing tasks and 62% on stacking, generalizing to objects never seen during training.

VR Teleoperation

Extend Robotics

Accessible VR teleoperation using ROS-native xArm to solve labor shortages and safety risks in hazardous environments.

VR Teleoperation & Autonomous Bin Picking

AMR Labs

AMR Labs developed VR teleoperation and autonomous bin picking applications with UFACTORY xArm.

Not sure which model fits your workflow?

Our team will match the right robot and configuration to your needs.