Skip to content
humanoidrobots.training

Sim benchmark · 31 models · policies · 0 submitted

Leaderboard

Full benchmark with method, sortable columns and per-model pages: bench.humanoidrobots.training

Two boards. The sim benchmark runs every published model headless each night under one generic controller, so a failure here is a fact about the model and our controller, not a verdict on the robot. The policy board ranks trained policies once they arrive with an eval we can rerun on the same pinned MJCF.

Nightly stand-hold benchmark

31/31 hold a stand · run 2026-10-11 · MuJoCo 3.14.0 WASM

Stand-hold benchmark per simulation model. Passing models first, then alphabetical.
RobotModel commitStand-hold 10 sBase driftRTFLoadDateMethod
AgiBot X1AgiBote6651b9agibot_x1_trainpassCoM min 98% · tilt 1.9°-0.9 cm16.7×2.25 s2026-10-11stand-hold-v2
Apptronik ApolloApptronikc96a32dmujoco_menageriepassCoM min 97% · tilt 1.6°-2.3 cm36.7×4.10 s2026-10-11stand-hold-v2
Berkeley Humanoid LiteUC Berkeley (Hybrid Robotics)fc90fedberkeley-humanoid-lite-assetspassCoM min 99% · tilt 0°-0.3 cm21.6×2.62 s2026-10-11stand-hold-v2
Booster K1Booster Robotics12a9751booster_assetspassCoM min 99% · tilt 0°-0.3 cm13.1×0.79 s2026-10-11stand-hold-v2
Booster T1Booster Roboticsc96a32dmujoco_menageriepassCoM min 99% · tilt 5.8°0.0 cm27.1×0.59 s2026-10-11stand-hold-v2
Booster T1 12-DoFBooster Roboticsda396a0booster_gympassCoM min 99% · tilt 0.3°-0.3 cm45.1×0.59 s2026-10-11stand-hold-v2
DropbearHyperspawndba8cd0dropbear-wbcpassCoM min 99% · tilt 2.2°-1.3 cm70.7×0.05 s2026-10-11stand-hold-v2
EngineAI PM01EngineAIc41043fengineai_ros2_workspacepassCoM min 98% · tilt 9.6°-0.9 cm18.0×3.87 s2026-10-11stand-hold-v2
Fourier GR1-T1Fourier48dc081Wiki-GRx-MJCFpassCoM min 99% · tilt 0.7°-0.4 cm18.4×0.98 s2026-10-11stand-hold-v2
Fourier GR2Fourier11da5a7Wiki-GRx-MJCFpassCoM min 99% · tilt 0.3°-0.3 cm23.2×0.61 s2026-10-11stand-hold-v2
Fourier N1Fourierc96a32dmujoco_menageriepassCoM min 100% · tilt 0.6°-0.1 cm19.7×3.20 s2026-10-11stand-hold-v2
K-Scale K-BotK-Scale Labs08defcekbot-modelspassCoM min 99% · tilt 2.2°-0.4 cm25.1×1.09 s2026-10-11stand-hold-v2
K-Scale Z-BotK-Scale Labsf51d6eakscale-assetspassCoM min 98% · tilt 0.1°-0.3 cm30.8×0.44 s2026-10-11stand-hold-v2
Leju Kuavo 4 ProLeju Roboticsc77dc2eLeju-GMRpassCoM min 100% · tilt 1.1°-0.2 cm14.6×1.23 s2026-10-11stand-hold-v2
Menlo AsimovMenlo Researchf71f3feasimov-1passCoM min 99% · tilt 0.3°-0.3 cm62.0×2.42 s2026-10-11stand-hold-v2
PAL KangarooPAL Robotics3bc51f8kangaroo_simulationpassCoM min 94% · tilt 1.3°-4.9 cm2.6×2.75 s2026-10-11stand-hold-v2
PAL TALOSPAL Roboticsc96a32dmujoco_menageriepassCoM min 99% · tilt 1.9°-0.3 cm4.5×0.89 s2026-10-11stand-hold-v2
PNDbotics Adam LitePNDboticsc96a32dmujoco_menageriepassCoM min 99% · tilt 1.5°-0.6 cm25.4×0.21 s2026-10-11stand-hold-v2
Pollen Reachy 2Pollen Roboticsf6d8284reachy2_mujocopassCoM min 100% · tilt 0°-0.3 cm10.0×0.93 s2026-10-11stand-hold-v2
Poppy HumanoidPoppy Projectd66c8fbpoppy-humanoidpassCoM min 98% · tilt 4.3°-0.3 cm18.2×0.08 s2026-10-11stand-hold-v2
Rainbow RB-Y1Rainbow Roboticsc96a32dmujoco_menageriepassCoM min 99% · tilt 0.1°-0.3 cm25.9×0.72 s2026-10-11stand-hold-v2
ROBOTIS OP3ROBOTISc96a32dmujoco_menageriepassCoM min 99% · tilt 0.8°-0.3 cm21.6×2.81 s2026-10-11stand-hold-v2
Roboto OriginRobopartyc71fd92roboto_originpassCoM min 100% · tilt 0.1°-0.3 cm5.2×2.08 s2026-10-11stand-hold-v2
ToddlerBot 2XCStanfordc96a32dmujoco_menageriepassCoM min 100% · tilt 1.6°0.0 cm9.7×3.43 s2026-10-11stand-hold-v2
ToddlerBot 2XMStanfordc96a32dmujoco_menageriepassCoM min 100% · tilt 1.6°0.0 cm10.1×2.98 s2026-10-11stand-hold-v2
Unitree G1Unitree Roboticsc96a32dmujoco_menageriepassCoM min 100% · tilt 0°+0.2 cm14.5×1.61 s2026-10-11stand-hold-v2
Unitree G1 12-DoFUnitree Robotics276801eunitree_rl_gympassCoM min 99% · tilt 0.4°-0.4 cm25.7×1.70 s2026-10-11stand-hold-v2
Unitree H1Unitree Roboticsc96a32dmujoco_menageriepassCoM min 100% · tilt 0.7°-0.4 cm23.6×0.95 s2026-10-11stand-hold-v2
Unitree H1 10-DoFUnitree Robotics276801eunitree_rl_gympassCoM min 99% · tilt 0.8°-1.1 cm29.9×1.03 s2026-10-11stand-hold-v2
Unitree H1-2Unitree Robotics1eb6642unitree_mujocopassCoM min 99% · tilt 0.7°-0.4 cm15.7×1.75 s2026-10-11stand-hold-v2
Unitree R1Unitree Robotics1eb6642unitree_mujocopassCoM min 99% · tilt 1°-0.3 cm17.1×1.67 s2026-10-11stand-hold-v2
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.
RTF
Simulated seconds divided by wall seconds for the 10 s run, single thread, no rendering. Depends on the machine; see environment.
Load
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.
Machine
local · 13th Gen Intel(R) Core(TM) i9-13980HX ×32 · win32 10.0.26300 x64 · Node v23.11.0
Data
Written nightly by the Coach agent (engine) to apps/humanoidrobots.training/data/benchmarks.json and merged through a pull request.

Policy board

  • Walk 0
  • Walk, rough terrain 0
  • Push recovery 0
  • Motion imitation (dance) 0
#PolicyRobotTrainerEval returnSim stepsRan on hardware

No policies yet. Zero is the real number.

Train one with mjlab and send the checkpoint, its eval video and the task id. The first entry goes here.

Set up a run