Menlo/asimov-locomotion-bench

Asimov Locomotion Benchmark

Can an Asimov-1 velocity-command locomotion policy stay on its feet? 65 short scenarios (walking across the trained command range, single pushes while standing and walking, sustained pushes, slippery floors, a 100 m lap and terrain: stairs, ramps, scattered boxes and rough tiles), each run several times with seeded actuator delay and joint noise. It is benchmarked through cyclotron, which simulates each run in MuJoCo with its humanoid-policy-viewer harness.

# in a cyclotron checkout (https://github.com/menloresearch/cyclotron)
./cyclotron.sh --benchmark <hub policy, e.g. org/name> datasets/Menlo/asimov-locomotion-bench --jobs 2
./cyclotron.sh --benchmark submit logs/benchmark_results/<timestamp>                   # upload the result to the policy's repo
./cyclotron.sh --benchmark pull datasets/Menlo/asimov-locomotion-bench                 # a local copy in benchmarks/, to edit
./cyclotron.sh --benchmark ./my-policy-folder benchmarks/asimov-locomotion-bench       # local policy or copy: can't be submitted

--benchmark only runs, locally. --benchmark submit uploads a result of a Hub policy on this Hub benchmark to that policy's .eval_results/Menlo__asimov-locomotion-bench.yaml (a direct commit if you can write to the repo, otherwise a pull request; --pr asks for one, --dry-run previews it).

Tests

One row per test in data/<category>/test.jsonl.

category	tests	what
locomotion	13	forward (0.6 and 0.8 m/s), backward, side-steps, diagonal, turns in place, figure of eight, squares at three speeds, the full command envelope, start/stop
push_standing	20	one 0.15 s push per test on a standing robot: chest, pelvis (front, back, side), each shoulder (front, outward, down), each at a lower and a higher force
push_walking	16	the same while walking forward at 0.6 m/s: chest, pelvis back, each shoulder front, outward and down
push_sustained	4	0.6 s pushes sideways at the pelvis and the left shoulder, at both forces
friction	4	straight walk and figure of eight on a wet (friction 0.3) and icy (0.08) floor; nominal is 0.6
distance	1	a 100 m square lap at 0.6 m/s, three minutes
terrain	7	walk straight at 0.4 m/s for 24 s over 5 cm and 10 cm stairs, 10° and 15° ramps, 3–7 cm boxes scattered across a 3 m lane, 0.5–4 cm rough tiles, and the 10° ramp with one 150 N sideways shoulder push

Every test has at most one push, so each push meets an undisturbed robot and an early fall cannot hide later pushes. Commands stay inside the range Asimov policies are trained on (vx −0.6…0.8 m/s, vy ±0.5 m/s, yaw rate ±0.8 rad/s), so no policy has its commands clamped. A push starts at least 1 s into its test and ends at least 2 s before it, so the robot has settled before and has time to recover after.

Terrain

Each course starts 1.5 m ahead of the robot and is built from the parameters in the test's terrain column (the harness generates the geometry, so every machine walks the same course). A fall on terrain is judged by the base's height above the ground under it. Every terrain test has an end zone: a 1 m circle 0.5 m past the end of the course (the test's goal column). A run only counts as surviving if the robot also reaches it, so stopping in front of a ramp is not a pass. 24 s at 0.4 m/s leaves room to finish at about 85% of the commanded speed.

Push forces

Every push location has two tests, named by their force (e.g. push_standing/chest_front_215n and chest_front_355n). The forces come from a sweep on a reference policy, model_Aug_18_1 (the checkpoint bundled with humanoid-policy-viewer), which pushed each location at 13 forces with the same 5 seeded repeats as the benchmark to find F50, the force where its survival first drops below one half:

• the lower force is 0.7 × F50: the reference survives with margin;
• the higher force is 1.15 × F50: the reference mostly falls, leaving headroom for better policies.

Both forces sit in the same category; there are no difficulty tiers. The forces are relative to that policy in this simulator, not a property of the real robot.

Scoring

benchmark.yaml runs every test 5 times (the distance lap twice). Each repeat draws its own actuator delay from the policy's env.yaml range (0–100 ms for Asimov) and up to 0.02 rad of initial joint noise, from a seed derived from the test and repeat, so runs are reproducible and independent of how they were scheduled.

A run falls when the base drops below 0.35 m (above the ground under it), the torso tilts past 78.5°, or the robot tumbles. On a test with an end zone, a run that never reached it counts as not surviving.

leaderboard	metric
survival_rate_v1	survival_rate@2	probability of ending a test upright (and in its end zone, if it has one); a robot that falls and gets back up survives
survival_time_v1	survival_time@2	1 for a run that ends upright, otherwise the fraction of the test before it last went down
<category>_survival_rate_v1	survival_rate@2	the same, for one category (locomotion, push_standing, push_walking, push_sustained, friction, distance, terrain)

Repeats of a test are averaged, tests are averaged within their category, and categories count equally. Each result carries a 95% interval over the seeds in its notes. Higher is better for all of them.

Reference results

Menlo/asimov1-locomotion-0818 (commit 18114a7), the same network as model_Aug_18_1, the reference policy above; all 322 runs, through cyclotron with humanoid-policy-viewer 55f7a07:

leaderboard	value	95% interval
survival_rate_v1	0.6559	0.5598–0.7519
survival_time_v1	0.8505	0.7564–0.9446
locomotion_survival_rate_v1	1	0.9149–1
push_standing_survival_rate_v1	0.5	0.4314–0.5686
push_walking_survival_rate_v1	0.5625	0.4827–0.6423
push_sustained_survival_rate_v1	0.65	0.4785–0.8215
friction_survival_rate_v1	0.85	0.6785–1 (falls on the icy straight walk, 3 of 5)
distance_survival_rate_v1	1	0.3999–1
terrain_survival_rate_v1	0.0286	0–0.1499 (falls on the stairs and boxes; stays up on the ramps but never reaches their end zones)

By construction the reference survives the lower push forces and mostly falls at the higher ones, so it scores about 0.5 on each push category; a better policy shows up as surviving the higher forces.

Versions

Task ids end in the benchmark version, so changing tests, repeats or scoring starts new leaderboards instead of mixing incomparable numbers. Each result also records the exact cyclotron and harness commits, benchmark protocol, dataset revision and policy commit that produced it.