Robotics · Shuja Abidi

Results

What ran, and what we measured

Four robots and several ways of moving, in the MuJoCo and PyBullet simulators. None of these used new RL training: Pepper needs no RL for this task, and the Go2, G1 and Tencent MAX run policies that others have trained and published.

RobotMethodMain result
PepperMediaPipe + arm IK, no RLCopied a person's arms from one video, with a 1.2° average forearm direction error against MediaPipe's estimate
Go2 quadrupedTrained RL policyReached 3.85 m/s when asked for 4, climbed 1.84 m of stairs, and stayed up through 200 N side pushes
G1, dancesOne trained policy per danceAll 28 dances finished without a fall, with a median joint error of 0.082 rad
G1, GentleHumanoidOne force-limited policy, many motionsFollowed new motions with no retraining, and its force limit changed how far a push moved the hand
Tencent MAX vs Go2Dog motion capture retargeted; pretrained MAX policiesOn one forest course, the MAX jumped 25 cm logs and the Go2 used its lidar to go round them; both crossed wet ground without falling
Everything here is from simulation, and nothing ran on a real robot. The errors compare the robot with its own reference motion, so they show how well it follows the plan, not how closely it matches a real person. See Limits.

1

Pepper: MediaPipe + arm IK

A person's arm motion from an ordinary video, copied onto a simulated Pepper.

Simulated Pepper seen from the side with one forearm reaching forward
7 sForearms reaching forward
Simulated Pepper from the front, with elbows bent and both forearms forward
8.5 sBoth forearms held out in front
Simulated Pepper from an angle, with one arm extending forward
13 sArms extending forward

Robot only. The test video belongs to the GentleHumanoid authors, so the person isn't shown. The robot uses simple shapes built from Pepper's published joint data (Apache 2.0), not its real appearance.

How it works

MediaPipe finds the person's shoulders, elbows and wrists. The direction of each arm segment is measured relative to the person's chest, and damped inverse kinematics finds Pepper's four joint angles per arm that point the same way. The motion is then played in physics, with gravity and Pepper's published motor limits. The steps are explained on Mimic Human.

Measurements, on one 16-second video

MeasureResultWhat it means
Person found100% of framesMediaPipe never lost them in this clip
Upper arm direction error3.9° average, 14° at the 95th percentilePepper against MediaPipe's estimate
Forearm direction error1.2° average, 4.3° at the 95th percentilePepper against MediaPipe's estimate
Tracking in physics0.034 radSimulated joints against their targets
Peak motor torque83% of the published limitWithin limits, with no joint saturated
Top joint speed2 rad/sCapped on purpose, well below Pepper's limits
Centre of massAt least 8 cm inside the wheelsA calculation; the base was fixed in place

What we learned

2

Go2 quadruped: a trained RL policy

A published walking policy, tested with no retraining.

Go2 running on flat ground
SpeedAsked for 4 m/s, reached 3.85 m/s
Go2 climbing a flight of grey steps
StairsUp 1.84 m and back down, with no steering or camera
Go2 staying upright during a side push
PushWalking on through a 150 N side push

How it works

The policy is go2_rl_lab (Wu et al., RSS 2026, MIT licence), trained in simulation on mixed terrain. Fifty times a second it reads the last 10 frames of body turning rate, tilt, speed command, joint angles, joint speeds and previous actions: 450 numbers in all. It outputs 12 joint target angles. It has no camera, so it senses stairs only through its legs.

Measurements, against an earlier flat-ground policy

Testgo2_rl_labEarlier flat-ground policy
Asked for 1 m/s0.91 m/s0.98 m/s
Asked for 2 m/s1.91 m/s1.85 m/s
Asked for 4 m/s3.85 m/s2.58 m/s
Backward, asked for 2 m/s1.95 m/s1.37 m/s
Turning at 3 rad/s2.99 rad/sFell at 1.7 s
Side pushes, 0.2 s, two tries eachNo falls up to 200 N; 1 of 2 at 300 N1 of 2 at 150 N; 2 of 2 at 300 N
Stairs and slopesClimbed all three test coursesFell on two, stuck on the third

This comparison favours go2_rl_lab: the earlier policy was trained only on flat ground and on a different robot model, and both ran here on the go2_rl_lab authors' model and scenes. At low speed, the earlier policy tracks speed slightly better.

A real gait change

No one programmed a change of gait. As the speed rises, a flight phase simply appears:

Actual speed0.91 m/s1.91 m/s2.95 m/s3.85 m/s
All four feet off the ground0%20%36%41%

What we learned

3

Humanoid G1: dance policies

One trained policy per dance: very accurate, but it can do only that dance.

G1 humanoid in a dance pose with one arm raised overhead
DiscoShort dance, 17 s
G1 humanoid dancing, with knees bent and arms out
HypeThe fastest arm motion tested
G1 humanoid dancing in a low pose
TikTokA lower pose, with knees bent
G1 humanoid dancing with an arm extended
SpiralAn arm swung wide

How it works

The G1 Moves dataset (CC BY 4.0) publishes a BeyondMimic-style policy for each motion. Each policy is rewarded for matching a single recorded dance while staying balanced, so it can specialise completely. Each one reads the target pose for the current moment, and the robot's state, then outputs 29 joint targets at 50 Hz.

The dataset's own runner made the robot fall within a second, because its settings didn't match the policies. The correct default pose, action scale, gains and body reference were inside each policy file, and using a matching robot model fixed it. As the page on math to AI says, settings have to match training.

Measurements

MeasureResult, over 28 dances and 19.8 minutes of motion
Finished without a fall28 of 28
Joint errorMedian 0.082 rad; range 0.066 to 0.152 rad
Body position errorMedian 4.8 cm; worst dance 13.9 cm on average
Peak motor torqueMedian 81% of limit, and some dances reach the limit

4

Humanoid G1: GentleHumanoid

One policy for many motions, which gives way when something pushes on it.

G1 humanoid at the start of a hug, with arms beginning to lift
Hug: startArms begin to rise
G1 humanoid midway through a hug with arms forward
Hug: reachingArms come forward
G1 humanoid at the end of a hug motion, with arms held forward
Hug: closingIn free air, with no person there

How it works

GentleHumanoid (Lu, Feng et al., Stanford) is trained on many human motions, while simulated forces push and pull on its arms. It tracks a spring-damper reference with a cap on the driving force, and you can set that limit from 5 to 15 N while it runs. It doesn't need a force sensor at run time: it learned how to respond in training.

The force limit changes behaviour

G1 humanoid with its right arm pulled far out by a push, at a 5 newton limit
Limit 5 NThe same 20 N pull moves the arm a long way
G1 humanoid with its right arm moved a little by a push, at a 15 newton limit
Limit 15 NThe arm resists more and moves less
Force limitHand moved by a 20 N pushHug arm error, in free airFalls
5 N36.6 cm0.162 radNone
10 N22.6 cm0.134 radNone
15 N15.5 cm0.122 radNone

Softer costs accuracy. As the limit falls, the hand gives way further, but the arm also follows the hug less closely, even with nothing touching it.

General policy or dedicated policy?

We ran both on the same reference motion, in free air, for each dance:

Arm joint errorGentleHumanoidDedicated dance policy
Median of 27 dances0.199 rad0.081 rad
Needs training for a new motionNoYes
Force limitYes, 5–15 NNone set
Falls in 28 dances1, at 52.6 s of a long dance0

The trade-off: the dedicated policy is about 2.5 times more accurate on arms. GentleHumanoid can follow new motions and give way to contact. Which is better depends on the task: a stage dance favours the first, and contact with people favours the second.

The gap grew with motion speed. It wasn't always in the dedicated policy's favour: on one slow dance, GentleHumanoid was the more accurate. The hug has no dedicated policy yet, so its gap can't be measured directly.

5

Lifelike dog: Tencent MAX and Unitree Go2

Working with the data and policies from Tencent Robotics X’s “Lifelike Agility and Play”.

61 seconds: motion-capture retargeting, then the MAX and the Go2 on the same forest course with logs, gaps and wet ground.

How it works

Lifelike Agility and Play (Han et al., Nature Machine Intelligence 2024) publishes dog motion capture and pretrained policies for Tencent’s MAX robot. We used both:

Measurements

CourseTencent MAXUnitree Go2
Five logs, 8–14 cm8.5 s, no fall11.5 s, no fall (no sensors used)
All logs one heightPassed 15, 20 and 25 cm; 30 cm only after strugglingPassed 15 cm; fell at 20 and 25 cm
Four 25 cm logs with gapsJumped all four, 9.1 sFound each gap with its own lidar and A*, 13.4 s
Gap course with mud, slick and ice stripsNo falls in 3 runs; one run stuck about 3 s at the last logNo fall, 63 foot slips, 14.7 s

The two robots ran in different simulators (PyBullet and MuJoCo), so times aren’t a strict comparison. The MAX’s results varied from run to run.

6

Limits of these results

What these numbers can't tell you.