01 / RECOVER
Down, but not out.
After a fall, a dedicated recovery policy brings the hand back onto its fingertips.
21/25successful hardware recoveries
84% across two fall directions
84% across two fall directions
ROBOTICS, WITH A LITTLE INDEPENDENCE
Fingers
as Legs.
Learning Self-Supported Locomotion and Manipulation
with an Anthropomorphic Hand
MEET THE HANDSELF-CONTAINED · 818 G
Off the arm.
Into the world.
Onboard power. Onboard computation. All five fingers.
Into the world.
01 THE IDEA
Same fingers.
New job description.
An off-the-shelf robotic hand learns to walk, support itself and interact with the world using the fingers it already has.
5fingers doing double duty
20powered joints
818 gincluding power & computation
14surfaces demonstrated
01 / THE PLATFORMThe original finger design and built-in joint controller stay. A compact module adds power, sensing, and a computer.
02 / THE CHALLENGELift one finger to step or press a key, and the others must still support the whole hand.
03 / THE APPROACHThe hand learns through trial and error in a calibrated simulator, then runs the learned policy onboard.
THE RESEARCH FILM
See it for yourself.
From the first steps to self-supported interaction.
Watch on YouTube ↗02 OUT OF THE LAB
Taking the scenic route.
Untethered crawling, demonstrated on 14 indoor and outdoor surfaces. No wheels. No extra legs. Just a hand getting on with it.
ALSO DEMONSTRATED ON
Rubber mat · Carpet · Tile · Metal grating · Asphalt · Dry concrete · Cut stone · Artificial turf
03 MORE THAN A WALK
A surprisingly
handy skill set.
The hand can also recover after a fall, press keys and push objects while supporting its own weight.
02 / INTERACT
A hands-on gamer.
Successive key presses execute operator-issued Sokoban moves, without visual feedback.
29/32correct keyboard commands
0.25 s median correct-press latency
0.25 s median correct-press latency
03 / MANIPULATE
A little push goes a long way.
With overhead visual feedback, the hand approaches and pushes a cube to a target.
17mmmean final target error
across 15 shown cube deliveries
across 15 shown cube deliveries
DEMO SETUP
Each skill uses its own policy. Keyboard trials start with manual alignment. Steering remains asymmetric.
04 HOW THE HAND LEARNS
Train in simulation.
Take the skill outside.
The actor learns to move through trial and error. A critic with extra simulator information helps train it.
ASYMMETRIC ACTOR–CRITIC PPOThe actor moves.
The actor moves.
The critic helps it learn.
CRAWLING · SIMULATION
AVAILABLE ON THE ROBOT
Sense + command
Joint angles, tilt, rotation rate and previous outputs, plus a movement command.
ACTOR · THE POLICY
Choose the movement
Turns sensor history and commands into 20 joint-target adjustments.
CONTROLLER + HAND
Try it in simulation
The built-in controller moves the simulated hand toward those targets.
New readings feed the next action.EXTRA HELP WHILE LEARNINGSimulation only
A WIDER VIEW
More information
Body state, joint torques and fingertip contact forces from the simulator.
CRITIC · THE ESTIMATOR
Assess the situation
Estimates the future reward expected from the current state.
VALUE ESTIMATE
How promising is it?
Helps PPO compare outcomes with expectations.
Collected experience
+ rewards + critic estimates
PPOTHE LEARNING ALGORITHM
Update actor + criticThen collect another batch.
Asymmetric: the critic gets extra simulator information during training.
1. Sense
Eight recent observations give the actor a short history of the hand’s motion.
Only the actor is deployed. The critic and PPO stay in training.
Inside the networks Inputs & training
The crawling actor, in numbers.
Eight samples of 49 inputs form a 392-value observation. A feedforward network with hidden layers of 512, 256 and 128 units produces 20 joint-target adjustments.
Learning happens in batches.
Training uses 4,096 parallel environments. Proximal Policy Optimization (PPO) uses a clipped objective to discourage overly large policy updates.
Architecture and training in the paper, Secs. III–IVSo, what makes
a useful step?
Rewards guide the learning. Explore one of them: a virtual-spring preference for each fingertip.
A HANDS-ON EXPLANATIONThree ideas. Try them yourself.
One finger, up close
Reference position Fingertip∿ Virtual spring
See the reference move with the body.
Start with
one fingertip.
The ring marks this fingertip’s reference in the settled stance. Each finger has its own.
This reference moves with the body.
Room to step.
A reason to stay close.
Stretching a virtual spring adds a training cost. Forward–backward movement costs less than sideways or upward movement.
At its referenceFarther away
Training cost0.36
Forward: a softer virtual spring.
Relative units. Compare directions at the same distance.We set the goals.
It learns the moves.
WE GUIDE
Reference positions, movement speed and stepping pace.
THE POLICY LEARNS
Which finger lifts when, and how far it moves.
Compare an earlier and a later lift at the same pace.
Soft guidance. Learned coordination.Illustrative geometry and motion.
For the technically curious Reward, reference frame & timing
The footprint objective
Each fingertip has a fixed reference relative to the body in a frame calibrated to the settled stance. This frame removes the nominal palm tilt, then rotates with the body.
Cost per finger = 0.25 × forward² + sideways² + vertical²
Displacements are measured from each fingertip’s reference. The five costs enter the reward with weight −30.
A pace, not a prescribed sequence
The lift objective encourages a pace that increases with commanded speed. Direction shaping discourages airborne motion against the command.
Read the formulation in the paper, Sec. IVSIMULATION · 12 TRAINING SEEDS
A better fit for this hand.
In simulation, our formulation produced faster movement than tuned rewards for four-legged robots.
Proposed formulation1.67 cm/s
Tuned quadruped rewards1.02 cm/s
Difference: +0.65 cm/s · 95% CI [+0.26, +1.02]
256 randomized episodes per seed. See paper, Sec. VI-A.
05 READ THE RESEARCH
All the details.
At your fingertips.
Fingers as Legs: Learning Self-Supported Locomotion and Manipulation with an Anthropomorphic Hand
VIEW THE PAPER Abstract
A walking robotic hand must use the same fingers to move its body, support its weight, and interact with the environment. We show how an anthropomorphic hand can learn these skills while retaining its finger design and position controller. Onboard power and computation make the platform self-contained. Our reinforcement learning approach accounts for the hand's unequal fingers, with training in a simulator calibrated from hardware measurements. In simulation, the hand moves faster with our reward formulation than with tuned rewards originally designed for quadrupeds. On hardware, task-specific policies enable untethered crawling, steering, and fall recovery. While supporting its own weight, the hand also executes successive keyboard commands without vision and pushes an object to targets using overhead visual feedback. These results demonstrate a compact mobile manipulator that reuses its fingers for locomotion and interaction, without a separate locomotion mechanism.
Citation
@misc{kazemipour2026fingers,
title = {Fingers as Legs: Learning Self-Supported Locomotion
and Manipulation with an Anthropomorphic Hand},
author = {Kazemipour, Amirhossein and Zheng, Hehui
and Katzschmann, Robert},
year = {2026},
eprint = {2609.17172},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.17172}
}