01 · Manipulation
Arms + vertical lift
A shared bimanual workspace across working heights.
Both arms use one action space while the lift repositions that workspace vertically, including while the base is moving.
HALE 1.0
We build the bodyand the brain.
A 22-degree-of-freedom bimanual mobile robot on actuators we design ourselves, run by a model pretrained on 150,000 hours of video.
Hardware
The robot architecture, built together.
Four systems form one body. OS3 owns the body architecture, custom actuators, control stack, model and training loop.
02 · Mobility
Base
Lateral positioning in tight rooms.
Holonomic drive moves the whole workspace sideways without turning the robot in place.
03 · Perception
Head + cameras
Independent gaze stays on the active workspace.
The head keeps the scene cameras aimed while the arms, lift and base continue moving.
04 · Actuation
Custom actuators
The mechanical layer OS3 builds in house.
One actuator architecture runs across the upper body and exposes full-state commands to the shared control stack.
- Arms
- 6+1 DoF
- Lift
- 180 lb
- Base
- 440 lb
- Actuators
- 22 total
- Cameras
- RGB + depth
- LiDAR
- Front + rear
- Battery
- Full workday
- Compute
- No GPU onboard needed
9 lb nominal, 15 lb peak, per arm.
Telescoping torso, floor to top shelf.
Omnidirectional. Fits standard doorways.
Quasi-direct-drive, backdrivable, torque-controlled.
Head cameras standard. Chest and wrist depth on developer units.
Front LiDAR standard. Rear LiDAR on developer units.
Charges from a standard outlet.
Cloud planner. Local manipulation policy.
Designed and built in San Francisco, California.
The model
One planner. One executor.
Long-horizon reasoning and whole-body action run at different timescales, then meet in one control path.
System 01
Planner, in the cloud
Maintains task context across long workflows. It stays outside the servo loop, so planning latency does not set actuator timing.
System 02
Executor, on the robot
Uses one whole-body policy with a contact-rich manipulation mode. Switching modes does not hand the task to a separate subsystem.
Control stack
From tokens to torque
Owns the timing boundary between policy output and hardware, keeping streamed motion smooth and compliant.
Read the control-stack articleTeleoperation and data
Four ways to show the robot a task.
Demonstrations come from people, with or without a robot in the room. A new task starts from two to four hours of them.
Rig 01
VR teleoperation
Drive both arms and the base from a headset, with the robot's own cameras as your eyes.
Rig 02
3D-printable rig
A printable handheld controller that mirrors the arm kinematics and can be reproduced wherever demonstrations are collected.
Rig 03
Exoskeleton
Wearable arms that track the operator's joints directly for high-fidelity bimanual demonstrations.
Rig 04
Handheld gripper
A gripper with cameras and no robot attached. Collect first-person demonstrations anywhere the work happens.
The training loop
Failure becomes training data.
Production robots fail. The question is whether they finish anyway, and whether the next one fails less. Our loop runs on our own hardware, with failures from real floors.
- 01Run
Run on real work
The robot attempts the workflow and retries recoverable misses on the spot.
- 02Correct
Correct the failure
When retry is not enough, a remote operator completes the step and captures the correction.
- 03Improve
Train on failures
Reinforcement learning on OS3 hardware targets those failures for future policy updates.
Safety
Layered stops.
Built to work next to people, in rooms that were never designed for robots.
- Motion starts only when a person starts a task from the rear screen or the app.
- Collision avoidance from cameras and LiDAR.
- Emergency stop and automatic power cutoff.
- Watchdogs halt the base when commands go stale.
- Compliant arms that yield on contact.
- A stopped robot stays stopped until someone checks it.
Go deeper
From tokens to torque.
How the physics layer makes the metal keep up with the model: identified dynamics, feedforward-first impedance control, trajectory shaping, and the measurements to prove it.
Partner with OS3
Build on the full platform.
Start from the robot, model and training loop we already operate.