Skip to content
humanoidrobots.training

bench.humanoidrobots.training · stand-hold-v2

Does the model stand?

Before you train a policy on a robot model, the model has to load and hold still. Every published MJCF in the browser sim is compiled headless in MuJoCo 3.14.0 (the same WebAssembly build the tab runs) and held for 10 s under one generic controller. A fall here is a fact about the model plus our controller, not a verdict on the robot.

Models

31

Hold a stand

31/31

Fell

0

Run

2026-10-11

Runner

32 cores · local

Results are 41 days old; the nightly run has not landed since.

Results

Sort by any column · default: passing first · RTF depends on the machine · results.json

Stand-hold benchmark results. Column headers sort the table.
AgiBot X1agibot_x1pass98.5%98.5%-0.9 cm1.9°16.7×2.25 s36.1 kg12
Apptronik Apolloapptronik_apollopass96.9%97.5%-2.3 cm1.6°36.7×4.10 s80.9 kg32
Berkeley Humanoid Liteberkeley_humanoid_litepass99.0%99.2%-0.4 cm0.0°21.6×2.62 s16.3 kg22
Booster K1booster_k1pass99.3%99.4%-0.3 cm0.0°13.1×0.79 s19.7 kg22
Booster T1booster_t1pass99.1%99.5%-0.0 cm5.8°27.1×0.59 s31.6 kg23
Booster T1 12-DoFbooster_t1_gympass99.2%99.4%-0.4 cm0.3°45.1×0.59 s31.6 kg12
Dropbeardropbearpass98.6%98.7%-1.3 cm2.2°70.7×0.05 s56.2 kg22
EngineAI PM01engineai_pm01pass98.5%98.7%-0.9 cm9.6°18.0×3.87 s40.9 kg24
Fourier GR1-T1fourier_gr1t1pass99.4%99.6%-0.4 cm0.7°18.4×0.98 s53.5 kg32
Fourier GR2fourier_gr2pass99.4%99.5%-0.3 cm0.3°23.2×0.61 s39.7 kg23
Fourier N1fourier_n1pass99.8%99.8%-0.1 cm0.6°19.7×3.20 s39.7 kg23
K-Scale K-Botkscale_kbotpass99.4%99.5%-0.4 cm2.2°25.1×1.09 s36.7 kg20
K-Scale Z-Botkscale_zbotpass98.0%98.5%-0.4 cm0.1°30.8×0.44 s3.4 kg20
Leju Kuavo 4 Proleju_kuavo_s45pass99.6%99.7%-0.2 cm1.1°14.6×1.23 s51.7 kg28
Menlo Asimovmenlo_asimovpass99.3%99.4%-0.3 cm0.3°62.0×2.42 s32.2 kg23
PAL Kangaroopal_kangaroopass94.2%94.9%-4.9 cm1.3°2.6×2.75 s60.2 kg22
PAL TALOSpal_talospass98.8%99.4%-0.4 cm1.9°4.5×0.89 s94.0 kg32
PNDbotics Adam Litepndbotics_adam_litepass98.7%99.3%-0.6 cm1.5°25.4×0.21 s58.2 kg25
Pollen Reachy 2pollen_reachy2pass99.7%100.3%-0.3 cm0.0°10.0×0.93 s51.4 kg25
Poppy Humanoidpoppy_humanoidpass97.7%99.4%-0.3 cm4.3°18.2×0.08 s2.6 kg25
Rainbow RB-Y1rainbow_rby1pass99.1%99.4%-0.3 cm0.1°25.9×0.72 s128.7 kg26
ROBOTIS OP3robotis_op3pass98.8%98.9%-0.3 cm0.8°21.6×2.81 s3.1 kg20
Roboto Originroboparty_roboto_originpass99.5%99.5%-0.3 cm0.1°5.2×2.08 s33.8 kg23
ToddlerBot 2XCtoddlerbot_2xcpass99.5%99.8%-0.0 cm1.6°9.7×3.43 s3.5 kg30
ToddlerBot 2XMtoddlerbot_2xmpass99.7%99.8%-0.1 cm1.6°10.1×2.98 s3.8 kg30
Unitree G1unitree_g1pass100.0%100.2%0.2 cm0.0°14.5×1.61 s33.3 kg29
Unitree G1 12-DoFunitree_g1_12dofpass99.3%99.5%-0.4 cm0.4°25.7×1.70 s32.1 kg12
Unitree H1unitree_h1pass99.6%99.6%-0.4 cm0.7°23.6×0.95 s51.4 kg19
Unitree H1 10-DoFunitree_h1_10dofpass99.0%99.0%-1.1 cm0.8°29.9×1.03 s51.6 kg10
Unitree H1-2unitree_h1_2pass99.5%99.6%-0.4 cm0.7°15.7×1.75 s67.0 kg27
Unitree R1unitree_r1pass99.3%99.4%-0.4 cm1.0°17.1×1.67 s28.9 kg29

Method · stand-hold-v2

Controller
The browser sim's hold controller (packages/sim/src/engine.ts), mirrored headless: joint servos toward the stand keyframe (torque motors run as implicit-damped servos clamped to their own torque limits, with gravity feed-forward), plus an ankle-strategy balance term from torso tilt and, where the manifest sets it, CoM-over-feet feedback; per-model gains from the @hs/sim manifest. Models published without actuators get a position servo per joint and models without a stand keyframe get one computed with the feet on the floor (each change is listed per model). No policy, no harness, no pushes.
Pass
Runs the full 10 s without a physics reset, the centre of mass never drops below 80% of its start height, and it ends within 10% of the start. Base drift is reported alongside.
Fell
The centre of mass drops below 60% of its start height at any point, or the physics diverges (NaN / auto-reset). Heights use the whole-body centre of mass because some models put the base origin near the floor.
Real-time factor
Simulated seconds divided by wall seconds for the 10 s run, single thread, no rendering. Depends on the machine; see environment.
Load time
Compile time: writing the pinned files into the WASM filesystem, mesh simplification when the model's runtime says so, MjModel.from_xml_path, and controller setup. Download time is excluded.
Models
Pinned to a commit of the model's own repo (see each model page). Some large meshes are simplified to fit the WASM heap, flagged per model.

Environment

MuJoCo
3.14.0 (WASM)
Node
v23.11.0
Platform
win32 10.0.26300 · x64
CPU
13th Gen Intel(R) Core(TM) i9-13980HX
Cores
32
Runner
local
Generated
2026-10-11T14:49:12.777Z

Reproduce

From the Hyperspawn suite monorepo (Coach operator agent, nightly via GitHub Actions):

pnpm --filter @hs/engine engine coach
pnpm --filter @hs/engine engine coach --only unitree_g1,booster_t1 --seconds 10

Model files are cached in engine/.cache/coach/models (about 700 MB). Output: this page's JSON plus a dated report under engine/reports/coach.

Next boards

Stand-hold is the floor. Policy evals (velocity tracking error and falls under the same pinned MJCF) rank on the leaderboard once policies arrive with an eval we can rerun; the in-tab walk tests are on each policy page.