Overview
Dr Hexapus is a wearable backpack carrying two robotic arms - a proof of concept for human augmentation with supernumerary limbs: extra effectors that add manipulation capability alongside your own arms rather than replacing them.
It has gone through two generations. The first was built to be worn and demonstrated - light enough to carry around a conference for an afternoon. The second is a heavier, reconfigurable research platform for asking a harder question: can people actually collaborate with autonomous limbs mounted behind them, out of sight?
Developed with Luca Mattioli and Etienne Burdet in the Human Robotics group at Imperial College London.
The question
Most supernumerary-limb demonstrations share three limitations: the task is quasi-static, the limbs are teleoperated, and the user watches them the whole time. That is not how extra arms would be used. A back-mounted limb spends much of its time outside your field of view, shares your workspace, and is often mechanically coupled to you through whatever you are both holding.
So the interesting questions are not about the robot on its own:
- Can people adapt quickly to autonomous limbs they can't see, and rely on them?
- How does the way a limb moves - its speed, its path - affect how fluently a person works with it?
- When there are several limbs, can their coordination help the wearer's balance instead of upsetting it?
Generation 1 - the ICRA demo
- Two Unitree Z1 arms - 6 DoF, 4.3 kg and 2 kg payload each
- Aluminium tubular frame with shoulder straps and a hip belt, around 10 kg all-in
- Untethered - portable power on the frame
- Three control modes: direct joystick teleoperation, pre-trained motion sequences, and hand tracking through an Ultraleap camera so your own hands drive the extra ones
It was worn continuously for three hours at ICRA 2024 in Yokohama, running interactive demonstrations with attendees. The two videos above are this configuration - a lot lighter than what came next. The point was to show that powerful robotics can be genuinely wearable - not bolted to a table with a person standing in front of it.
Why the arms changed
The Z1 is light, but its base joint cannot hold the full arm moment when extended horizontally, and a 74 cm reach is too short to work alongside human arms rather than merely near them. Most other collaborative robots weigh 16-20 kg each, which rules them out on a back.
The Kinova Gen3 sits in the gap: 7 DoF, a reach of 89 cm that exceeds a human arm, 8 kg with a 4 kg payload. Back-mounted, that gives a shared workspace that overlaps the wearer's own reach in front and extends behind them to regions they cannot reach at all.
Generation 2 - the research platform
- Up to four Kinova Gen3 arms, two facing up and two facing down, with Robotiq 2F-85 grippers
- Rigid X-shaped frame on an industrial load-bearing vest that spreads the weight across shoulders and hips
- Reconfigurable mounts - each arm base clamps to a rail through a turntable and a sliding element, giving continuous adjustment of position and rotation. Moving an arm from the upper rails to the lower ones takes under five minutes with a single key.
- Heavy - roughly 30 kg with two arms. Comfort is the honest limitation of this generation, and the price of using commercial cobots with real payload.
- Emergency stops within reach of the wearer at all times
The platform lives inside Imperial's Multi-limb Virtual Environment (MUVE): Vicon motion capture, a split treadmill with a force plate under each foot, immersive VR, and up to four arms, all time-synchronised under ROS. That combination is what makes the questions above measurable - you can see the wearer, the limbs and the object, and read the ground reaction forces while they work together.
What we can study with it
A representative experiment: an assembly task that cannot be done with two arms. A long tube has to be picked up on one side of the body, passed over the head, carried to the other side, and screwed to a support - which needs the tube held steady and hands free for the screws at the same time. Holding a three-metre tube in one hand is not an option; the wrist torque alone rules it out.
The limbs work autonomously through it:
- Task state, not joystick - the sequence is a state machine. The wearer's right hand approaching the tube triggers the first limb; each following step is triggered by the previous one completing, or by where the wearer's hands and head are.
- Goals computed online for each wearer's body and posture, so grasp poses and the over-the-head handover are placed correctly without offline tuning.
- Safety-aware planning - trajectories come from optimisation with collision avoidance against a voxelised scene, including a virtual wall behind the wearer's back that no limb may cross. No training data involved, which keeps the motion legible and the guarantees inspectable.
- Biomechanics-aware co-manipulation - when a limb and a hand hold the same object, the limb's path is shaped so the human arm follows a comfortable route, not one that forces awkward shoulder rotation.
- Coordination between limbs - a second limb can mirror the working one as a counterweight, so moving eight kilograms of robot to one side does not tip the wearer.
Kinematics, per-foot ground reaction forces, and questionnaires then tell you how the person adapted, how fluent the collaboration was, and what it did to their balance. A study along these lines is in preparation.
Engineering analysis
The repository includes an analysis suite for the frame, switchable between the Z1 and Gen3 configurations: mass properties and centre of gravity, static FEA under overhead and asymmetric loads, modal analysis, stability and tipping envelopes, and an ergonomic loading assessment.
Student work on this platform
Two MSc projects at Imperial build on Dr Hexapus this year - one on predicting the wearer's intent from vision so a limb can assist before being asked, the other on decoding shoulder movement from sensors in the backpack straps to drive a limb continuously.
Applications
Anywhere a task needs more hands than a person has, and a second person is impractical: assembly and maintenance work, surgical assistance holding instruments steady, and assistive use for people whose own limbs are limited.
From marble catchers to extra arms
The first project on this site is a tool an eight-year-old built to reach marbles down a drain. This one gives people extra arms. The instinct is the same.