bench.humanoidrobots.training · stand-hold-v1

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

Hold a stand

11/15

Fell

Run

2026-09-24

Runner

32 cores · local

Results are 13 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.

Apptronik Apolloapptronik_apollo	pass	98.5%	98.8%	-1.1 cm	1.5°	55.5×	2.98 s	80.9 kg	32
Berkeley Humanoidberkeley_humanoid	pass	97.9%	97.9%	-1.2 cm	4.5°	49.0×	1.30 s	16.1 kg	12
Booster T1booster_t1	pass	99.1%	99.5%	-0.0 cm	4.9°	29.5×	0.41 s	31.6 kg	23
Booster T1 12-DoFbooster_t1_gym	pass	94.8%	95.6%	-2.7 cm	0.3°	51.8×	0.39 s	31.6 kg	12
Fourier N1fourier_n1	pass	99.8%	99.8%	-0.1 cm	0.7°	38.2×	2.69 s	39.7 kg	23
PAL TALOSpal_talos	pass	98.8%	99.4%	-0.4 cm	1.9°	6.0×	0.72 s	94.0 kg	32
ROBOTIS OP3robotis_op3	pass	92.5%	92.9%	-2.1 cm	0.6°	33.6×	1.87 s	3.1 kg	20
Unitree G1unitree_g1	pass	100.0%	100.2%	0.2 cm	0.0°	26.4×	1.09 s	33.3 kg	29
Unitree G1 12-DoFunitree_g1_12dof	pass	99.6%	99.7%	-0.2 cm	0.8°	42.2×	1.14 s	32.1 kg	12
Unitree H1unitree_h1	pass	99.6%	99.6%	-0.4 cm	0.7°	41.9×	0.56 s	51.4 kg	19
Unitree H1 10-DoFunitree_h1_10dof	pass	98.8%	98.8%	-1.4 cm	0.8°	56.0×	0.71 s	51.6 kg	10
Agility Cassieagility_cassie	fell	17.0%	17.0%	-84.8 cm	88.3°	8.9×	0.46 s	33.3 kg	10
Dropbeardropbear · meshes simplified	fell	13.3%	13.3%	23.3 cm	103.0°	0.6×	8.14 s	61.3 kg	19
PNDbotics Adam Litepndbotics_adam_lite	fell	11.2%	11.4%	-83.0 cm	90.1°	55.1×	4.13 s	58.2 kg	25
ToddlerBot 2XCtoddlerbot_2xc	fell	32.4%	32.5%	-19.8 cm	149.8°	15.9×	2.52 s	3.5 kg	30

Method · stand-hold-v1

Controller: The browser sim's hold controller (packages/sim/src/engine.ts), mirrored headless: joint PD toward the stand keyframe (or qpos0) plus an ankle-strategy balance term from torso tilt, per-model gains from the @hs/sim manifest. 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 10.0.26200 · x64
CPU: 13th Gen Intel(R) Core(TM) i9-13980HX
Cores: 32
Runner: local
Generated: 2026-09-24T18:09:18.839Z

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.