Leaderboard

Lunar robotics campus with three demonstration stations and a central first, second and third place podium.

First results

11 November

π0.5
X-VLA
G0.5
Awomo
DeepSeek V4.1 Flash
Gemini 4 Argon

Leaderboard

VLA / WAM Leaderboard

VLA / WAM Leaderboard
Rank	Model	Type	Contributor					Results
—	π0.5	VLA	Physical Intelligence	—	—	—	—	11 Nov
—	X-VLA	VLA	Tsinghua AIR	—	—	—	—	11 Nov
—	G0.5	VLA	Galaxea AI	—	—	—	—	11 Nov
—	Awomo	WAM	Awomo	—	—	—	—	11 Nov
—	GR00T N1.7	VLA	NVIDIA	—	—	—	—	Queue #1
—	StarVLA	VLA	StarVLA	—	—	—	—	Queue #2
—	GigaWorld-Policy	WAM	GigaAI	—	—	—	—	Queue #3
—	DM0.5	VLA	Dexmal	—	—	—	—	Queue #4
—	OpenWAM	WAM	OpenWAM	—	—	—	—	Queue #5
—	KinRT	VLA	IIGroup, Tsinghua	—	—	—	—	Queue #6
—	EWAM	WAM	EWAM Team	—	—	—	—	Queue #7
—	X-WAM	WAM	Tsinghua & Xiaomi	—	—	—	—	Queue #8
—	Liber-0	WAM	LiberAI	—	—	—	—	Queue #9
—	InternW0	WAM	Shanghai AI Lab	—	—	—	—	Queue #10

Coding Agent Leaderboard

Coding Agent Leaderboard
Rank	Model	Type	Contributor					Results
—	DeepSeek V4.1 Flash	Coding agent	DeepSeek	—	—	—	—	11 Nov
—	Gemini 4 Argon	Coding agent	Google	—	—	—	—	11 Nov

Coding agents run a 5-case subset of every task × difficulty (450 per setting); end-to-end models run all 50 (4,500). Because case counts and interfaces differ, the two boards are ranked separately and their scores should not be compared. We plan to extend agents to the full setting as resources allow.

Evaluation schedule

11 November First results

π0.5
X-VLA
G0.5
Awomo
DeepSeek V4.1 Flash
Gemini 4 Argon

Then After 11 November, results for about one more model every two days. In order:

GR00T N1.7
StarVLA
GigaWorld-Policy
DM0.5
OpenWAM
KinRT
EWAM
X-WAM
Liber-0
InternW0

Reading the leaderboard

Full evaluation protocol →

Score

A successful case scores 100. Otherwise the task's partial-credit rule applies, capped by difficulty: 30 / 22 / 15 for Mind and Act tasks and 15 / 10 / 7 for Mind2Act tasks (Easy / Medium / Hard). Four tasks give no partial credit. The board reports the mean over valid cases, zeros included.

Success %

The share of valid cases that scored 100. Score and Success are separate measurements; a high score with low success means many partial completions.

Overall and suites

Overall is the plain mean across all 30 tasks; Mind, Act and Mind2Act are the means over their ten tasks each. With 50 cases per task × difficulty, averaging over cases, cells or tasks gives the same number.

Clean and Random

Clean uses the original scenes. Random changes only appearance and distraction: 23 table materials, 6 light types with varied colour and intensity, and table clutter. Objects, robot, cameras, instructions and scoring are identical, and every Random case pairs with a Clean case. The two are reported separately.

What each board measures

VLA / WAM

Cases: 30 tasks × 3 difficulties × 50 = 4,500 per setting
Interface: Three 640×480 RGB views, joint state and the instruction in; 14-D joint actions out at 30 Hz
Budget: Physics steps = 1.2 × the P90 expert duration of each task × difficulty; no call limit

Coding agents

Cases: 5 of the 50 cases per task × difficulty = 450 per setting
Interface: RoboProbe tools: end-effector targets planned by cuRobo, wait, screenshot memory and give up
Budget: 4 × the end-to-end physics budget and at most 150 decisions per case

Different case counts, interfaces and budgets make the two boards incomparable, so each is ranked on its own. As resources allow, we will extend coding agents to the full 50-case setting.

How results are produced

The evaluation set is hidden and has no validation split; evaluation seeds are separate from training seeds.
Mind2Act runs every model on its own machines with the released code, assets and budgets. Each case is run once; no uncertainty is reported.
Infrastructure failures are retried once and then excluded; valid zeros are kept. Each result reports its valid cases out of the total.
Submitters declare the model, version, checkpoint SHA and inference settings (agents: model, effort, transport, call limit and retries).