Humanoid · legged robots

A 4-legged robot starts on flat ground and walks over an uphill slope, a downhill slope, rough terrain and 10 stair steps to the landing at the top of the stairs. Every section spans the full path width, so none can be bypassed.

7 of 7 verdict items passed · 2 not measured · same verdict across 2 repeated runs of the same test

Scenario's own summary: Pass — not scored by the judge; the overall verdict takes the worse of this and the items above.

• Walking is done by an AI locomotion policy that imSim trained on the rough-terrain locomotion task of a public robot learning environment. The route (7 waypoints) is set by the scenario, and velocity commands come from the imSim waypoint follower.
• A gentle test course: uphill and downhill slopes of about 5° (height 0.7 m), rough terrain of 0.5 m boards 2 to 8 cm high, and stairs with 5 cm steps, 10 steps (0.5 m).
• The course has no obstacles or people, so common items such as near miss and emergency stop read 0 without any stimulus. Collision and anomalies are not measured in this test.
• Fall, waypoint and tilt criteria are set by this scenario. 30° is the start angle of the imSim fall warning range. The other criteria are imSim common defaults used in all domains.
• The tilt verdict value excludes the tilt at the start (about 1.7°). The maximum of 10.5° occurred while walking the course, with large tilts on the uphill, rough terrain and stair sections. Relative to gravity, the maximum is about 12°.
• Both runs used the same settings. The two runs moved through the course sections identically and had the same maximum tilt (10.5°). Only the motion after reaching the end point differed slightly. This is not a robustness test with varied conditions.
• Ground friction (1.0) and foot friction (0.8, 0.6 when sliding) match the training settings.
• The robot is a 4-legged robot imported from a public model. These are not humanoid results.
• These are simulation results and do not replace physical testing.

The same terrain course completion test is run once per step, changing only the robot's total body mass (×1 to ×3), to see up to which mass it walks to the end without falling or tilting heavily.

Pass up to ×1 · Fail from ×1.5

1 of 6 stages passed · one run per stage

• The masses of the robot body and its 12 leg links were scaled by the same factor (feet and rotational inertia unchanged). This differs from loading cargo on the body.
• Joint force limits are unchanged, and the AI locomotion policy that handles walking is not aware of the mass change. imSim trained the policy on the rough-terrain locomotion task of a public robot learning environment.
• ×1.5 passed all 7 waypoints without falling. It swayed heavily sideways on the rough terrain (34°), exceeding the tilt threshold (30°), so the result is Fail.
• ×1.75 passed all 7 waypoints, then collapsed and flipped over while standing on the landing. It fell once, but the sensors recorded again while the flipped body rocked, so 4 falls were recorded.
• The tilt verdict value for ×1.75 (45.9°) is the value at the moment of falling. Tilt is not measured after a fall, so the flipped posture is not included in this value.
• From ×2 up, the robot could not support its body while settling before the start and collapsed. Tilt at these steps was measured relative to the already collapsed posture and is not used. Falls are 2 because two sensors both recorded the single collapse.
• Each step was run once. The boundary lies between ×1 and ×1.5.
• Verdict criteria are the values set by this scenario (falls, waypoints, tilt) and imSim common defaults. Tilt excludes the tilt at the start.
• Ground friction (1.0) and foot friction (0.8, 0.6 when sliding) match the training settings.
• The robot is a 4-legged robot imported from a public model. These are not humanoid results.
• These are simulation results and do not replace physical testing.