Motion Tracking Leaderboard

Reproduce this evaluation →

Build a comparison

Each metric uses its own scale; GPU Hours is logarithmic. Missing values are n/a.

Policies

MimicLite-ROAMimicLite-PPOHEFTHoloMotionSONICSONIC-Low-LatencySONIC-v1.1GRIT v0.0.1ScaleBFM-XLScaleBFM-MBFM-ZeroTeleopItHumanoid-GPTTWIST2

Metrics

Body PosBody OriGlobal RootGPU HoursWrist PosWrist OriTracking ReturnProgress

Datasets

MotionDecode

Locomotion (80)Manipulation (48)Ground (60)Dance (40)

Legacy

Legacy LAFAN-40Legacy PHUMA-30Legacy Root-90

Datasets

MotionDecode

Locomotion (80)Manipulation (48)Ground (60)Dance (40)

Legacy

Legacy LAFAN-40Legacy PHUMA-30Legacy Root-90

Loc = Locomotion, Man = Manipulation · Click any metric header to rank.

Rank	Policy
1	MimicLite-ROA	31.98Loc 29.22 · Man 34.74	0.106Loc 0.095 · Man 0.118	0.261Loc 0.261 · Man —	173.2—	33.21Loc 31.61 · Man 34.81	0.154Loc 0.141 · Man 0.166	1.826Loc 1.886 · Man 1.766	96.6Loc 98.7 · Man 94.5
2	MimicLite-PPO	32.92Loc 31.52 · Man 34.32	0.109Loc 0.097 · Man 0.121	0.255Loc 0.255 · Man —	92.3—	35.96Loc 33.25 · Man 38.66	0.135Loc 0.117 · Man 0.153	1.836Loc 1.900 · Man 1.772	97.6Loc 100.0 · Man 95.2
3	HEFT	32.98Loc 32.23 · Man 33.73	0.100Loc 0.102 · Man 0.098	0.403Loc 0.403 · Man —	116.0—	33.46Loc 30.60 · Man 36.31	0.109Loc 0.112 · Man 0.105	1.888Loc 1.897 · Man 1.879	99.4Loc 100.0 · Man 98.7
4	SONIC-Low-Latency	34.84Loc 30.73 · Man 38.95	0.164Loc 0.156 · Man 0.171	0.889Loc 0.889 · Man —	21000.0—	38.64Loc 33.24 · Man 44.05	0.206Loc 0.177 · Man 0.236	1.787Loc 1.812 · Man 1.761	100.0Loc 100.0 · Man 100.0
5	GRIT v0.0.1	35.08Loc 31.72 · Man 38.43	0.152Loc 0.147 · Man 0.158	0.395Loc 0.395 · Man —	——	33.09Loc 28.02 · Man 38.17	0.243Loc 0.238 · Man 0.247	1.817Loc 1.838 · Man 1.797	100.0Loc 100.0 · Man 100.0
6	SONIC-v1.1	35.11Loc 29.55 · Man 40.67	0.160Loc 0.158 · Man 0.161	0.623Loc 0.623 · Man —	21000.0—	38.33Loc 31.98 · Man 44.68	0.167Loc 0.163 · Man 0.172	1.797Loc 1.815 · Man 1.780	100.0Loc 100.0 · Man 100.0
7	HoloMotion	35.74Loc 32.97 · Man 38.51	0.158Loc 0.145 · Man 0.170	0.889Loc 0.889 · Man —	9200.0—	33.29Loc 30.19 · Man 36.38	0.231Loc 0.215 · Man 0.246	1.757Loc 1.795 · Man 1.718	98.5Loc 100.0 · Man 97.1
8	TeleopIt	36.23Loc 36.25 · Man 36.22	0.149Loc 0.146 · Man 0.152	0.340Loc 0.340 · Man —	400.0—	46.46Loc 46.65 · Man 46.27	0.209Loc 0.197 · Man 0.221	1.821Loc 1.831 · Man 1.811	100.0Loc 100.0 · Man 100.0
9	ScaleBFM-M	36.30Loc 36.56 · Man 36.03	0.131Loc 0.131 · Man 0.131	0.962Loc 0.962 · Man —	——	37.33Loc 38.96 · Man 35.69	0.200Loc 0.201 · Man 0.199	1.841Loc 1.841 · Man 1.840	100.0Loc 100.0 · Man 100.0
10	SONIC	37.10Loc 28.73 · Man 45.47	0.160Loc 0.151 · Man 0.170	0.363Loc 0.363 · Man —	21000.0—	43.45Loc 30.15 · Man 56.75	0.228Loc 0.211 · Man 0.245	1.773Loc 1.817 · Man 1.729	100.0Loc 100.0 · Man 100.0
11	ScaleBFM-XL	37.98Loc 38.21 · Man 37.75	0.133Loc 0.148 · Man 0.117	0.633Loc 0.633 · Man —	——	51.24Loc 54.15 · Man 48.32	0.166Loc 0.180 · Man 0.152	1.846Loc 1.826 · Man 1.865	100.0Loc 100.0 · Man 100.0
12	Humanoid-GPT	38.75Loc 42.71 · Man 34.78	0.141Loc 0.148 · Man 0.133	1.323Loc 1.323 · Man —	15000.0—	50.64Loc 53.12 · Man 48.16	0.194Loc 0.184 · Man 0.203	1.811Loc 1.786 · Man 1.836	99.7Loc 99.3 · Man 100.0
13	TWIST2	50.30Loc 56.84 · Man 43.77	0.340Loc 0.341 · Man 0.338	3.868Loc 3.868 · Man —	——	68.98Loc 75.32 · Man 62.64	0.649Loc 0.652 · Man 0.645	1.343Loc 1.300 · Man 1.387	96.6Loc 95.0 · Man 98.1
14	BFM-Zero	52.11Loc 56.68 · Man 47.54	0.242Loc 0.253 · Man 0.231	3.090Loc 3.090 · Man —	——	87.12Loc 89.87 · Man 84.38	0.355Loc 0.334 · Man 0.375	1.537Loc 1.498 · Man 1.575	98.4Loc 100.0 · Man 96.9

Metric scope

Tracking Return measures pelvis-XY/yaw-aligned local-pose quality. Progress and Tracking Return use the same pelvis-height/projected-gravity termination, and reward is counted only before termination.

Protocol

• The default held-out benchmark is MotionDecode Locomotion-80, Manipulation-48, Ground-60, and Dance-40. Legacy L = LAFAN-40, P = PHUMA-30, and R = Root-90 remain selectable.
• Each headline is the unweighted mean of the selected, available dataset means. Each dataset mean gives every motion equal weight; the second line shows those dataset means directly.
• Body Pos / Body Ori: local-body position/orientation error; lower is better.
• Global Root: root-only tracking error; lower is better. MotionDecode reports this only for Locomotion, so the other three MotionDecode sets are n/a and excluded from the headline.
• GPU Hours: training cost when recorded; populated values link to their authoritative source. Missing cells mean this audit did not find a defensible official source. The checked-in audit manifest is docs/src/data/gpu_hours_provenance.json.
• Wrist Pos / Wrist Ori: mean left/right wrist position/orientation error; lower is better. MotionDecode Ground and Dance are n/a and excluded from the headline.
• Tracking Return: normalized tracking return; higher is better.
• Progress: motion completion; higher is better.
• All policies use seed 0, initial_pause_s=0.0, no runtime cap, and the same failure conditions. See Cross-Codebase Tracking Evaluation for the full definition and reproduction command.

Last updated: September 2, 2026.