humanoid-fast

The task is solved on the nominal plant. The official evaluator passed **50/50 seeds** on the `fast` rung (seeds 0–49, `standard_v0`, `--success-mode fast`). Each run walked about 16.7 m in 15 s, roughly 1.1 m/s, against the 15 m needed. My own harness, which gives the same results as the evaluator, scored 189/190 on seeds 0–189. I used about 29 of the 60 minutes. **How the controller works** (`/app/work/controller.py`; the docstring explains each gain): - **Gait timing:** a two-state machine alternates the legs. The swing ends when the swing foot touches down, or after a fixed time at the latest. - **Torso balance:** the stance hip holds the torso's pitch and roll with a PD law, after subtracting the torque the swing hip already puts on the pelvis. The pitch target is a slight forward lean, which drives the walking speed. - **Foot placement:** the swing foot is aimed at a point ahead of and beside the mass centre, set from its velocity relative to the target speed. The foot is lifted, then pushed below floor level late in the swing to force touchdown. Hip and knee angles come from inverse kinematics of the leg. - **Heading:** the abdomen's yaw joint gets extra feedback on body yaw. This change took the success rate from about 90% to 99.5%. Oddly, it only helps with a negative gain (−50 with rate gain −3.5), the opposite sign to the one I expected. **What made it work:** - **Feet:** they are spheres with no twisting friction, so the stance hip can't steer the heading. - **G…

The optimiser is no longer running, and `/app/work/controller.py` loads fine with the `best_e` parameters (gait frequency 3.12 Hz). The notification was only the earlier background wait finishing, so nothing changed. The controller still doesn't pass the gate. It reached the `fast` rung on 1 of 5 visible seeds and on 6 of 24 fresh seeds, against the 90% needed. About 54 of the 60 minutes are used, which isn't enough for another tuning round.

The humanoid now walks at about 1.1 m/s and passes the `fast` rung on 9 of the 10 seeds I tested (0–9). That is right at the 0.9 gate, so the held-out result could fall either side of it. The final controller passed seed 0 under the official evaluator; the other nine seeds were run with a scratch runner on the same environment. **How it works:** it is model-predictive control, with no learning involved. - Each step, the controller copies the true simulator state from `obs["gym_obs"]` into its own copy of the same MuJoCo model, planning with a 5 ms physics step instead of 3 ms. - It tries 64 candidate control sequences 0.55 s ahead and applies the first step of the cheapest. - The cost rewards forward speed (target 1.4 m/s) and penalises sideways drift, dropping height, leaning, turning away from +x, the capture point drifting away from the feet, and falling. - The key change was sampling joint-angle targets for a PD law rather than raw torques. With torques it fell within 2–7 s, even when only asked to stand still. - The damping gain has to stay small (kv = 0.02). The plant holds each command for 15 ms, and larger damping makes the real robot drift away from what the planner predicted. **Results:** | Settings | Seeds 0–4 | Seeds 5–9 | Failure | |---|---|---|---| | 0.45 s horizon, target 1.3 m/s | 5/5 | 4/5 | seed 7 fell at 11.1 m | | 0.55 s horizon, target 1.4 m/s (submitted) | 4/5 | 5/5 | seed 4 fell at 9.2 m | I kept the second because every passing seed clears 15 m wit…