Hardware

Robot platforms

The hardware we build on. The 3D models below are rendered from the same URDF files the robots run on, and every platform listed is bookable by members unless its entry says otherwise.

Models are rendered from the URDF of each platform. Drag-free: move the pointer to look around.

Specifications

AI Worker

ROBOTIS FFW-SG2 · bimanual mobile manipulator

A bimanual mobile manipulator, and the platform our RoboCup@Home entry is built on. The swerve-drive base moves omnidirectionally, which is what lets it work in the tight spaces a furnished arena leaves between furniture.

It is the resource three activities compete for — the research track, the competition team and public demonstrations — so the research track's schedule is fixed first and everything else is planned around it.

Model: ROBOTIS AI Worker FFW-SG2
Degrees of freedom: 25 total, 7 per arm
Payload: 3.0 kg per arm nominal, 5.0 kg peak (6.0 / 10.0 kg dual)
Reach: 641 mm to wrist, plus hand
Mobile base: Swerve drive, omnidirectional
Sensing: 3 × RGB-D (ZED Mini head, 2 × RealSense D405 in the hands), 2 × LakiBeam 1 LiDAR, IMU
Onboard compute: NVIDIA Jetson AGX Orin 32 GB
Battery: 25 V, 80 Ah (2,040 Wh)
Size and weight: 604 × 602 × 1,623 mm, 90 kg
Units in the lab: 1
Access: Members, after platform-specific safety training

Manufacturer

OMY

Robot arms · manipulation policy experiments

Two robot arms used for manipulation policy experiments and data collection. Run as a leader–follower pair, they collect teleoperation data directly.

Degrees of freedom: 6
Payload: 3 kg
Reach: 580 mm
Repeatability: ±0.05 mm
End effector: RH-P12-RN gripper with Intel RealSense D405
Units in the lab: 2, run as a leader–follower pair
Access: Members, after platform-specific safety training

Manufacturer

LeRobot SO-ARM 101

Teaching platform · imitation learning

Two small robot arms used in the education track. A full cycle — collecting demonstrations, training a policy, evaluating it — runs short enough on these that a member new to robot learning can complete one in a session.

Degrees of freedom: 6 per arm
Actuators: Feetech STS3215 bus servos with magnetic absolute encoders
Configuration: Leader and follower arms
Units in the lab: 2
Access: Members; reserved for group practicals on Thursday evenings

Manufacturer

Dexterous hands

Tesollo · ROBOTIS · LEAP

Three multi-fingered hands used for grasping and in-hand manipulation research. They are booked separately from the arms, so a study can pair any hand with any platform. New policies are checked in simulation for range of motion before they run on the Tesollo hand, whose finger links are thin enough to be damaged by excessive torque or collision.

Manufacturer pages: Tesollo Delto Gripper 3F, ROBOTIS RH-P12-RN, LEAP Hand.

Tesollo Delto Gripper 3F: 12 DOF (4 per finger), 1,044 g
ROBOTIS RH-P12-RN: Adaptive 2-finger, 0–106 mm stroke, 5 kg payload, 170 N
LEAP Hand: 16 DOF, open-source, Dynamixel-driven
Units in the lab: 1 of each
Primary use: Dexterous grasping, in-hand manipulation, hand hardware experiments

Sensors and data capture

Cameras · MANUS data glove

Seven cameras for robot vision and dataset capture, and a data glove that tracks finger joints so a person's hand motion can drive any of the robot hands directly. Camera placement and lighting are kept identical between sessions — a manipulation dataset only holds together if they are — and the layout is recorded whenever it changes.

Cameras: 7
Data glove: MANUS, 1 unit
Primary use: Robot vision, teleoperation, manipulation dataset capture

Compute

NVIDIA RTX PRO 6000 workstations

Three GPU workstations for simulation, policy training and vision model experiments. Long training runs are moved to nights and weekends so they do not collide with daytime experiments on the robots.

Workstations: 3
GPU: NVIDIA RTX PRO 6000 (Blackwell)
GPU memory: 96 GB GDDR7 ECC each
Cores: 24,064 CUDA cores, 752 fifth-generation Tensor Cores
Primary use: Simulation, policy training, vision model experiments

Manufacturer

AI Sapiens

Humanoid · previously operated

A humanoid platform SHAPE operated during its first year. It is no longer in the lab, and it is listed here because the experience of running it is part of what the team knows.

Actuators: 23 Dynamixel-Q quasi-direct-drive actuators
Status: Previously operated; not currently in the lab
Primary use: Whole-body control and locomotion experiments

Manufacturer