Wheeled humanoids
Dual arms on a stable base
Two arms, a waist and a wheeled base. Strong at stations, cells and flat floors, where the robot replaces a person standing at a table. Our reference platform is the AgiBot G2.
Services
Everything between the crate and the job
Manufacturers ship capable hardware with a basic controller and an SDK. Getting that robot to pack a carton or cross a hospital ward is separate work. Our deployment software automates most of it, and our engineers do the rest in person.
Deployment software, with engineers alongside
The software automates the site model, the training data, the training and the tests for each site. Our team scopes the work, integrates the robot and supports it after go-live. If you already own a robot, we start wherever it is stuck.
01 Application software
Task software that runs on the real robot
We write the logic for the job: where to pick, where to place, what to do with a reject and how to recover from a failed grasp. Every arm move goes through one motion server, which checks the full path against the robot's own body and the fixtures before a joint moves. Navigation runs on mapped zones with speed limits.
We build on the manufacturer's interface, such as the AgiBot GDK on the G2 and the low-level command interface on the Unitree G1. Code tested in simulation runs on the robot without a rewrite.
- Task state machines with tested failure paths
- Motion planning with collision and clearance checks
- Mapping, localisation and navigation
- A safety layer with a control lease, a watchdog, an emergency stop and speed zones
02 Policy training
Policies trained on your task, with gates before they ship
Some skills are easier to learn than to program. We collect demonstrations by teleoperation or with a scripted generator, record synchronised camera, joint and command data, and train imitation policies such as ACT and behaviour cloning.
For locomotion and whole-body control we train with reinforcement learning in MuJoCo MJX, thousands of robots in parallel, then distil a teacher policy into a student that only uses the sensors the real robot has. Each stage has pass gates, and a policy reaches the robot only after it passes in simulation the tests it must pass on site.
- Teleoperation by keyboard, SpaceMouse or webcam
- Episode recording to HDF5 and LeRobot-style datasets
- Imitation learning: ACT and behaviour cloning
- Reinforcement learning with PPO, curricula and teacher-student distillation
03 On-site integration
In person, on your site
A robot is finished when it works in your building with your people. We come to site to survey and map the space, set out the cell and the fixtures, and set speed zones and safety limits with your health and safety team.
We commission the robot against acceptance tests agreed with you before the work starts, train your operators and hand over with the test results. When your products or your layout change, we retrain and recommission.
- Site survey, mapping and cell layout
- Commissioning against agreed acceptance tests
- Operator training and handover
- Retraining for new products and layout changes
04 Simulation assets
Simulation assets for your use case
Work with us to build the simulation assets for your use case: your site, your fixtures and products, and candidate robots running your tasks. You get reach, layout and cycle-time evidence and an architecture for the deployment before you buy hardware.
Work with usPlatforms
We work with robots you can buy
We are not tied to a manufacturer. We choose the platform for the task and the building, and our software and training pipeline move with us from one robot to the next.
Bipedal humanoids
Legs and hands for buildings made for people
Legs for thresholds, ramps and tight rooms, with dexterous hands for doors and handling. Our reference platform is the Unitree G1 with BrainCo Revo2 hands.
Quadrupeds
Four legs for rounds and rough ground
Legged platforms for inspection rounds, carrying and uneven floors. The same simulation, training and safety pipeline applies, with the platform chosen to suit the route.
How a project runs
Five stages, each with something you can check
Every stage ends with evidence: a report, a recorded run or a passed test. You decide at each one whether to carry on.
- 01Scope
We visit, walk the process with your team and agree the task and the acceptance tests.
- 02Simulate
We model the site and the robot, then test reach, clearances and cycle time.
- 03Train
We collect demonstrations, train the policies and gate them against the tests.
- 04Integrate
We install on site, commission against the acceptance tests and train your operators.
- 05Support
We watch the numbers, retrain when things change and extend to the next task.
Have a robot that is not doing its job yet?
Tell us what it is, where it works and what it should be doing. We will tell you what it needs.