MABEL
MABEL
56 functional degrees of freedom driven by 59 motors, almost entirely 3D-printed, assembled in one weekend. One coherent platform — not a stack of parts.
Drag to orbit, scroll to zoom, open it full-screen — then move the joints. Every slider's range comes straight from MABEL's URDF; the model articulates live.
Seven subsystems, head to wheels. Pick a tab — each opens to its real parts, open files, and key numbers.
Two open-source ORCA hands give MABEL real dexterity — five tendon-routed fingers each, 17 actuated DOF per hand, driven by Feetech serial servos.
Each hand is the open-source ORCA hand — five fingers, tendon-routed, anthropomorphic. Print it, string it, run it; the design is fully released.
Sixteen Feetech HLS3915 finger servos plus an HLS3930 wrist, all on one 1 MHz TTL serial chain — abduction, MCP, and PIP per finger.
Pinch, power, and tripod grasps; the same position commands drive the simulated and the real hand, so a grasp tuned in sim works on hardware.
Assign servo IDs, write registers, and read live load and temperature from the ORCA Hand Control Studio — no vendor tool. See .
Derived from the open-source OpenArm design and driven by Damiao quasi-direct-drive motors on CAN — low gearing means real torque control, gravity compensation, and compliance.
Two 7-DOF arms built on the open-source OpenArm design — a proven, printable, research-grade arm rather than a closed industrial unit.
Damiao QDD motors on CAN close a high-rate current loop, so commanded torques become real ones with very little friction to fight.
Low gearing makes the arms back-drivable — the foundation for gravity compensation and Cartesian impedance in the whole-body controller.
Link lengths and shoulder spacing are tuned to a person's, so teleoperation maps naturally and the embodiment gap stays small.
A 3-module delta swerve on REV SPARK controllers gives true holonomic motion — strafe through a doorway, rotate in place, hold a heading while driving.
Three independently steered-and-driven modules in a delta layout — full holonomic motion in any direction, decoupled from heading.
NEO / NEO550 brushless on REV SPARK controllers — driven through a reverse-engineered CAN stack, zero dependency on REV's tools. See .
Shortest-path steering and a cosine anti-slip choke keep the wheels from scrubbing; the math is shared with simulation.
A direction-aware tip-over model shrinks speed and braking limits as the lift raises the CoM — agile, but it won't tip.
A cascaded Z-lift adds 0.635 m of vertical travel; an RP2040 Pico runs closed-loop position-plus-velocity control over a brushed actuator.
Telescoping stages give 0.635 m of vertical range in a compact stow height — counter height to over a table.
An RP2040 Pico runs cascaded position + velocity control over a BTS7960 H-bridge and a quadrature encoder, with a host library and an iOS GUI. See .
Raising the lift raises the centre of mass, so the base's tip-safe limits tighten with height — modelled in whole-body control.
Matching the operator's eye height shrinks the embodiment gap and makes teleop feel like wearing the robot.
A pitching torso and a 3-DOF Dynamixel neck turn MABEL from an arm-on-a-cart into an active, expressive body that can lean, look, and gaze.
A Damiao-driven torso pitch extends low reach and lets the whole upper body lean into a task instead of over-stretching an arm.
A 3-DOF Dynamixel neck (yaw / pitch / roll) aims the head cameras — active perception and natural gaze for the operator.
Eye height, lean, neck range, and arm length are matched to a person, so a teleoperator feels at home immediately.
Torso and neck are degrees of freedom the whole-body controller coordinates alongside the base, lift, and arms.
Exteroception to see the world, proprioception to know itself, and two surfaces to show what it's thinking — all published as ordinary topics.
A 3-DOF pan / tilt / roll neck aims a stereo RGB-D camera — active perception that points the gaze where the task is, and feeds the Vision Pro stereo passthrough.
One camera in each wrist gives close-up, rectified vision exactly where the grasp happens — the detail the head can't always see over the hand.
An RPLIDAR-class scanner under the base publishes /scan for SLAM, localization, and obstacle avoidance — the backbone of navigation.
Joint encoders on every actuator and wheel encoders on the swerve base give /joint_states and /odom; an IMU stabilizes the base estimate. The robot always knows its own pose.
A 13″ chest touchscreen and animated LED-matrix eyes surface MABEL's state and intent — expressive HRI, so people read it before it acts.
Every sensor publishes onto the ROS 2 state bus (or its sim twin); consumers see topics, never the hardware. Over teleop the same feeds stream through the server.
Custom base and body PCBs tie every actuator together over a single low-latency local network — CAN for the motor buses, Ethernet for the high-bandwidth links — with one Jetson Thor running it all onboard.
Purpose-built boards for the base and the body replace a rat's nest of breakouts — clean power distribution and bus routing for the whole robot.
CAN carries the motor buses; Ethernet carries the high-bandwidth links. Keeping the local network unified cuts latency across the system.
A single Jetson Thor runs SLAM, navigation, perception, and whole-body control — no cloud, no tether. See Software.
One battery powers the whole platform; nothing trails behind it as it drives the room.
