Bodies do not match
A human pose sequence cannot be copied joint-for-joint onto G1. Link lengths, joint limits, and contact geometry change the motion.
Embodied motion research / 2026
Teach a humanoid
to find the net.
We started with a simple question: can one recorded penalty kick become one believable robot skill? MESSI takes that clip, retargets it to G1, and checks the result in MuJoCo before claiming anything more.
G1 / 29 DOF
TARGET: GOAL LINE
MI300X
ROCm
ROCm
01 / THE THESIS
Start small.
Make it real.
Sports video captures timing, balance, and follow-through that are hard to script by hand. Our approach is deliberately narrow: take one meaningful action, preserve its character through retargeting, and judge it by what happens to the ball in physics.
01
Why this task
A penalty kick is compact, legible, and easy to score. One motion, one ball, one outcome.
02
How we approached it
Use the dataset's supplied pose, map it onto G1, then refine the trajectory on MI300X instead of inventing motion from scratch.
03
What counts as success
No proxy metrics. The kick only works if the ball crosses the line in MuJoCo.
02 / THE DYNAMICS GAP
Video is a useful reference. It is not a control policy.
Motors have limits
A human strike says nothing about actuator saturation or how much impulse reaches the ball once the robot meets the ground.
Physics decides
Grass-like rolling friction removes the easy answer: a kick must carry the ball through the goal mouth, not merely make contact.
03 / PIPELINE
From source clip
to replay.
Read the motion
Uses SoccerKicks 3D pose outputs from a recorded penalty kick. No pose-estimation claim: the project parses the dataset's supplied annotations.
↗Retarget to G1
Maps the reference motion into a Unitree G1 joint-control trajectory, preserving the signature swing, arm movement, and trunk rotation.
↗Tune on MI300X
Uses PyTorch/ROCm batched refinement on AMD MI300X to improve the trajectory, with scoring-tune optimization as the next concrete step.
↗Verify in physics
Replays the result in MuJoCo with a real ball, goal geometry, and grass-like rolling friction. The simulation is the final arbiter.
↗04 / AMD IN THE LOOP
Use acceleration
where the search
actually is.
Before settling on the shipped Lab path, we explored using AI acceleration where the loop is expensive: batching power, aim, timing, and yaw candidates in PyTorch/ROCm, then replaying the best shot in full MuJoCo physics.
POSE SOURCE
SOCCERKICKS→AI SEARCH
ROCm / MI300X→FINAL CHECK
MUJOCO
SoccerKicks supplies the pose annotations. AI-guided candidates narrow the control search. Goal-line crossing remains the acceptance test.SOCCERKICKS→AI SEARCH
ROCm / MI300X→FINAL CHECK
MUJOCO
05 / EVIDENCE
See the retargeted
kick run.
Choose any camera to inspect the same latest neutral-recovery replay: G1 starts ready, strikes, then returns upright after the ball is away. It is a real MuJoCo execution artifact, not a concept render or storyboard.
- Source: SoccerKicks penalty motion
- Robot: Unitree G1 in MuJoCo
- Views: overview, contact, goal line, side
REPLAY / KICK CAMNEUTRAL → KICK → RECOVERY
05A / MOTION PROOF
One kick.
Four viewpoints.
Same verified rollout, cut into the angles that make contact, flight, and recovery legible.
06 / INTERACTIVE PHYSICS
Place the keeper.
Find the lane.
The Goalie Lab puts you in charge of the defense. Drag the keeper across the goal mouth, then watch a learned policy pick its lane, power, and tempo — verified by a real MuJoCo rollout rendered straight from the simulator.
GOAL MOUTH3.66 m
Open the interactive Goalie Lab →Real MuJoCo physics rendered server-side. No install: move the keeper and get a fresh replay back.
GOAL MOUTH / 3.66 MKEEPER SWEEP
-1.58 mOPEN THE LAB TO TAKE CONTROL+1.58 m
07 / WHERE THIS GOES
One kick is a
useful unit.
The narrow penalty task is the proof point, not the ceiling. If demonstration-to-physics transfer works here, it becomes a repeatable way to study many compact embodied skills.
SPORTS TRAINING
Compare a reference action with a robot execution inside the same legible task.
VIDEO-TO-SKILL
A compact testbed for turning annotated demonstrations into controllable embodied behavior.
ROBOTICS RESEARCH
A reproducible place to study retargeting, contact, and trajectory refinement together.
Next: expand the motion bank, tune aim continuously, add a moving goalkeeper, and test the learned skill beyond simulation.