hoverhand
A 30 cm humanoid robot that flies on fans in its feet and uses its hands in the air.
I started with Microban, an open-source humanoid from the Rhoban team. I put a 102 mm ducted fan in each foot and two 127 mm rotors on its back. Each forearm has a gripper. The robot weighs 1.47 kg.
A hand-written controller flies the robot. A small neural network adds a correction to the thrust of each rotor. I trained that network with reinforcement learning.
Simulation only · the hardware build is next
Flight
The robot starts on the floor with its motors off. It takes off, flies to four points and lands on a new spot. Its arms move and its grippers open and close during the flight.
Simulation replay at real speed · hand-written controller with the learned correction
Results
It holds still while its arms move
When an arm moves, the center of mass moves too, and the arm pushes back on the body. In a 27-second test, the robot reached out, swung its arms fast and used its grippers. The hand-written controller kept it within 1.5 cm of its hold point during the reaches. During the fast swings, it stayed within 0.88 cm.
The full 27-second test at real speed · hand-written controller only
It keeps flying when a fan gets weaker
At 6 seconds, the right foot fan drops to 80% of its thrust. The controllers get no signal about this. The hand-written controller alone loses control 5.5 seconds later. With the learned correction, the robot gets to three of the four points within 10 cm. Then it lands upright, 5.4 cm from its target.
Same route and same fault, side by side from 4 s · the fan with the red rim is the weak fan · the left clip holds its last frame after the crash
The correction has a cost. With all fans healthy, it tracks the route a little less closely: 5.2 cm RMS, against 4.4 cm without it. But it lands much closer to the spot: 1.1 cm, against 19.1 cm.
It learned this in training
The network trained on 4,096 simulated robots at the same time. Training took about 25 minutes on one GPU. In one of every four training flights, one rotor lost 10–25% of its thrust. The first saved versions crashed on the test route. From 17 million training steps on, every version landed.
Four saved versions fly the weak-fan route at the same time · played at 2×
Every result on this page comes from one run of one route in a physics simulator
Next
Where it can go
The robot is small. It flies for about four minutes, and its grippers hold 20–30 g. But the idea can grow: a robot that flies to a place that is hard to reach, lands on what is there and uses its hands.
Disaster sites
A small robot can fly over a gap and land on debris. Each leg finds its own footing. Then it turns off its fans and works as a camera or a radio relay.
Power lines and towers
A larger version can land on a tower with its legs. It stops its fans and works with two hands. That uses almost no power.
Space
Fans do not work in a vacuum, so a space version must use cold-gas thrusters. The hard part stays the same: steer a body with moving limbs, and stay in control when a thruster fails.
- Simulation
- Earth flight
- Air-bearing table
- Reduced gravity
- Cabin, on fans
- Vacuum, on thrusters
Steps 03 to 06 are years away · the first step there is software: run this controller on an existing free-flyer testbed