walker2d-robust
Walker2d (MuJoCo Walker2d-v5) · solved 9 of 9 agent attempts pass
more
SB3-zoo TD3 (4701 nominal) falls on every seed with -20 % motor gain; the SIMBICON design from walker2d-forward stays alive on 40/40 seeds under every perturbation (2026-10-23 matrix) and is the oracle.
Gate. per-seed success is the `forward` rung of this benchmark's ladder ['alive', 'forward', 'brisk', 'fast'] (`--success-mode forward`), not the evaluator's default flag; the task is solved when the success rate over the grader's held-out seeds is `>= 0.9`; the same gate must also hold on the sibling scenario(s) `heavy_torso_20`, `low_friction_30`, `weak_motors_20`, `gravity_high_10` (`--scenarios`)
Axes: plant:walker2d, scenario:standard_v0+physics-suite, gate:forward_3m, robustness:torso-friction-motors-gravity
Each video re-runs the attempt's final controller through the task's own evaluator on a public seed (never the grader's held-out ones). The table under each video is the official grade on the held-out seeds.
standard_v0| criterion | graded | needed |
|---|---|---|
| base:forward | 1.000 | ≥ 0.900 |
| scenario:heavy_torso_20 | 1.000 | ≥ 0.900 |
| scenario:low_friction_30 | 1.000 | ≥ 0.900 |
| scenario:weak_motors_20 | 1.000 | ≥ 0.900 |
| scenario:gravity_high_10 | 1.000 | ≥ 0.900 |
The agent's final message
The controller at `/app/work/controller.py` passes the full gate with a wide margin, so I'm stopping here with about 28 minutes of budget unused. **Result.** The official evaluator reports solved on seeds 0-9 for the nominal plant and all four sibling scenarios (heavy torso, low friction, weak motors, high gravity), with every rollout reaching the brisk rung. On my own check over seeds 0-119 across all five graded plants, 600 of 600 rollouts stayed alive and walked at least 11.7 m, against a 3 m gate. | Metric (seeds 0-9, nominal) | Value | |---|---| | Ladder | alive 1.00, forward 1.00, brisk 1.00, fast 0.00 | | Final x (mean) | 12.7 m | | Return (mean) | 2592 | | Worst min torso height / max pitch (120 seeds) | 0.96 m / 0.52 rad | **Control law.** A two-state SIMBICON-style gait with PD joint tracking. The swing hip tracks a world-frame thigh angle with speed feedback for foot placement, the swing knee follows a smooth flexion bump for clearance, and the ankles hold the feet level in the world frame. The stance hip holds torso pitch and cancels the swing hip's reaction torque. Gains were found by a parallel random search over seeds 0-5 of all five plants. **Three findings drove the design.** The hip and knee ranges are extension-only, so the legs can only swing ahead of the torso when the torso leans back, and the gait holds a nose-up pitch of about 0.43 rad. The torso weighs only 3.7 kg, so the torso pitch loop must be soft or the commands chatter at the control rate. S…
A SIMBICON-style walker alternates right/left stance states by timer or swing-foot touchdown, using PD control of the swing thigh world angle with speed feedback, a sinusoidal knee flexion, stance torso-pitch PD with swing-reaction compensation, and ankle PD with a late-stance offset.
Shown: public seed 0 (standard_v0); this episode passes the forward rung. This episode: episode_steps 1000, x_final 12.7, min_height 1, max_abs_pitch 0.499, n_steps 31, total_reward 2591.
controller.py · agent transcript · trial walker2d-robust__vwam5jo
heavy_torso_20| criterion | graded | needed |
|---|---|---|
| base:forward | 1.000 | ≥ 0.900 |
| scenario:heavy_torso_20 | 1.000 | ≥ 0.900 |
| scenario:low_friction_30 | 1.000 | ≥ 0.900 |
| scenario:weak_motors_20 | 1.000 | ≥ 0.900 |
| scenario:gravity_high_10 | 1.000 | ≥ 0.900 |
The agent's final message
The controller at `/app/work/controller.py` passes the full gate with a wide margin, so I'm stopping here with about 28 minutes of budget unused. **Result.** The official evaluator reports solved on seeds 0-9 for the nominal plant and all four sibling scenarios (heavy torso, low friction, weak motors, high gravity), with every rollout reaching the brisk rung. On my own check over seeds 0-119 across all five graded plants, 600 of 600 rollouts stayed alive and walked at least 11.7 m, against a 3 m gate. | Metric (seeds 0-9, nominal) | Value | |---|---| | Ladder | alive 1.00, forward 1.00, brisk 1.00, fast 0.00 | | Final x (mean) | 12.7 m | | Return (mean) | 2592 | | Worst min torso height / max pitch (120 seeds) | 0.96 m / 0.52 rad | **Control law.** A two-state SIMBICON-style gait with PD joint tracking. The swing hip tracks a world-frame thigh angle with speed feedback for foot placement, the swing knee follows a smooth flexion bump for clearance, and the ankles hold the feet level in the world frame. The stance hip holds torso pitch and cancels the swing hip's reaction torque. Gains were found by a parallel random search over seeds 0-5 of all five plants. **Three findings drove the design.** The hip and knee ranges are extension-only, so the legs can only swing ahead of the torso when the torso leans back, and the gait holds a nose-up pitch of about 0.43 rad. The torso weighs only 3.7 kg, so the torso pitch loop must be soft or the commands chatter at the control rate. S…
A SIMBICON-style walker alternates right/left stance states by timer or swing-foot touchdown, using PD control of the swing thigh world angle with speed feedback, a sinusoidal knee flexion, stance torso-pitch PD with swing-reaction compensation, and ankle PD with a late-stance offset.
Shown: public seed 0 (heavy_torso_20); this episode passes the forward rung. This episode: episode_steps 1000, x_final 12.8, min_height 1, max_abs_pitch 0.5, n_steps 30, total_reward 2604.
controller.py · agent transcript · trial walker2d-robust__vwam5jo