humanoid-parity

The final controller is in place at `/app/work/controller.py` and passes the gate with margin. I'm stopping here with about 15 minutes of budget left, since the submission is already passing and the remaining candidates I screened were less robust. **Result** | Check | Value | |---|---| | Fast rung, 20 visible seeds (official evaluator) | 19/20 | | Fast rung, 196 fresh seeds (own harness) | 193/196 (98.5%) | | Return mean, base scenario | about 6300 (gate 5020) | | Mean speed | 1.16 m/s, about 17.5 m per episode | **Controller design** - A SIMBICON-style two-state walker. The stance leg alternates when the swing foot touches down, with time bounds on the step. - The swing foot is placed in world coordinates from the mass-centre position and velocity relative to the stance foot, both sagittally and laterally. This foot placement is what balances the robot, since the spherical feet give no ankle torque. - The stance hip regulates torso pitch, with partial compensation of the swing-hip reaction torque. Roll is handled the same way on the hip-roll motors. - All 26 gains were found with a hand-written CMA-ES in `scratch/cma.py` on random seeds, then every fully surviving candidate was re-screened on 40 to 100 fresh seeds before installation. No neural networks or gradients were used. **Dead ends and caveats** - An open-loop sinusoidal gait with pitch feedback could not stand for more than a second. Early SIMBICON gains saturated the 0.4 command range and the stance roll torq…

The controller does not solve the task. The `fast` rung needs ≥15 m, and the best I got is about 3.7 m per episode, so the success rate on seeds 0–4 is 0/5. I stopped at about 53 of the 60 minutes. `/app/work/controller.py` holds the best version I had. On the visible seeds 0–4 on the nominal plant: - **Ladder:** alive 5/5 and forward (≥3 m) 5/5. Brisk (≥8 m) and fast (≥15 m) are 0/5. - **Reward:** the mean is 5109, which clears the 5020 reward gate. That gate is separate from the success rate, and the success rate is what fails. - **Held-out check:** on 12 other seeds, mean reward was 5102, alive in all 12, with distances of 3.5–3.8 m. The controller is a clock-driven gait, with Fourier-series targets for the legs and each leg half a cycle behind the other. It adds PD joint torques and feedback from pitch, roll, yaw and forward and lateral speed. The numbers are in `THETA` and were found with a CMA-ES search I wrote (`opt.py`, `run1.py`). The search objective was reward plus a bonus for distance. The docstring in `controller.py` explains the control law and what each gain does. The gait is mostly a low-speed shuffle with both feet often off the floor, about 0.25 m/s, not a real stride. Search runs that used a stronger forward bonus or a wider search from this solution reached 5–6 m, but they were still improving slowly and had not reached 15 m when I stopped. I did not run any of the perturbation scenarios (low friction, weak motors, pushes and so on). One earlier solut…

I'm leaving the controller in `/app/work/controller.py`. It passes the grading gate on every seed I ran: 11 out of 11 reached the `fast` rung (stayed up for all 1000 steps and covered at least 15 m). **Results on the standard scenario** - **Official verifier, seeds 0–4:** 5/5 `fast`, mean return 6116 (the gate is 5020), mean distance 17.3 m, mean speed 1.15 m/s. - **My own rollout script, seeds 5–10:** all six survived the full episode, covering 16.2–18.1 m with returns between 5993 and 6171. The closest seed walked 16.2 m, so the margin over the 15 m bar is about 1 m. I only tested the standard scenario, not the perturbation scenarios. **How it works** It doesn't use a hand-designed gait. It uses a sampling-based model-predictive controller: - **Internal model:** the controller loads the same humanoid model file into its own MuJoCo copy. Each step it rebuilds the full simulator state from `gym_obs`, which contains every coordinate except horizontal position, and that doesn't matter on a flat floor. - **Plan:** it keeps a smooth command plan 0.6 s long (40 control steps, with a knot every 60 ms). Each step it tries 128 noisy variants of the plan, plus the current plan itself. It runs all of them forward in parallel in a cheaper planning model (7.5 ms timestep, simpler integrator), keeps the lowest-cost one and applies its first command. - **Cost terms:** forward speed of the mass centre towards 1.7 m/s, lateral drift, head height, staying upright, keeping the mass centre's…