hopper-blackout-noisy

`/app/work/controller.py` passes the gate on the blackout scenario: 98% `forward` success (59 of 60 seeds, 20–79) and 100% on seeds 0–19 and 0–4. The one failure was seed 54, which crashed at 8.4 m when the torso dropped below 0.7 m. The runs used the sandbox evaluator, `blackout_noisy_v1` and `--success-mode forward`, and about 46 of the 60 minutes were used. - **Policy:** it is a linear feedback law on the 11 Hopper-v5 observables plus a foot-contact flag. I found its gains with gradient-free random search, with no neural network. The first version stood still or crept. It only learned to hop once I dropped the +1 alive bonus from the search reward. - **Speed cap:** the final fine-tune capped the rewarded forward speed at 1.6 m/s, which cut falls. The gait still covers about 17.5 m per episode, which is `brisk` but not `fast`. - **Blackouts:** the controller runs a MuJoCo copy of the plant inside it, which assumes the grader's environment has `mujoco` and `hl_control` installed (as the sandbox does). It steps that copy with the last applied command, then checks each reading against the previous raw reading. A reading that repeats the previous one to within the noise, and fits that better than the prediction, is treated as stale and replaced by the prediction. Live readings are blended in, or adopted outright after a big disagreement such as an impact. - **Why the estimator matters:** the first, simpler estimator accepted too many stale readings, and the policy passed only …

Final state: `/app/work/controller.py` holds the best controller found. It does not reach the 0.9 gate on the blackout plant, so the result should be read as a partial solution. **Outcome on the task plant (`blackout_noisy_v1`), forward rung** | Seeds | Evaluator | Forward rate | |---|---|---| | 0-29 | official `verify` | 0.77 (alive 0.83) | | 0-69 | my rollout script | 57/70 = 0.81 | | 50-69 (never used for tuning) | my rollout script | 16/20 = 0.80 | | 0-9 on `standard_v0` | official `verify` | 1.00 (alive 1.00) | Failures split into two kinds: a pitch excursion past 0.2 rad on a landing (the robot falls), and episodes where the bounce dies out and the robot stands still short of 3 m. **What the controller does** - **Stance:** the hip is a pitch servo whose target leans back as speed exceeds the desired value. The knee is a stiff PD strut. The ankle acts as a leg spring that plantarflexes in proportion to how far the thigh trails behind the torso, so the toe push launches the forward-leaning body up and forward. An extra hip gain engages when pitch passes 0.15 rad. - **Flight:** PD to a landing posture, with the ankle set to land the foot flat and a weak hip pitch servo that uses the heavy leg as a reaction wheel. Knee and ankle commands ramp in over the first 0.4 s. - **Blackouts:** stale readings are detected when the joint angles stop moving as the measured rates say. For the following 9 steps the law runs on dead-reckoned angles and pitch, with height propagated ba…

The budget is nearly spent (50 of 60 minutes used), so I'm stopping here with the best controller saved at `/app/work/controller.py`. **Outcome: below the gate.** On the visible seeds under `blackout_noisy_v1` the submitted controller passes the forward rung on 4 of 5 seeds (0.8 success rate, gate 0.9), and on 9 of 12 seeds when I widened the range. The failure mode is always the same: a fall from pitching out of band, not a sensing error. | Visible-seed metric | Value | |---|---| | forward success (seeds 0-4) | 0.80 | | forward success (seeds 0-11) | 0.75 | | mean distance when alive | about 9 m | | mean return | about 1875 | **What the controller does.** It is a phase machine driven mostly by the clock: lean the standing robot forward with the ankle, then knee thrust plus toe push, a knee tuck to lift the foot, a flight phase with a hip PD mixing torso pitch and landing thigh angle, and a knee re-extension before touchdown. Thrust and toe push shrink when the robot is faster than the target speed. A stale-sensor filter detects blackouts from the collapse of the per-step change in the six measured rates, dead-reckons the state from the last live sample, predicts touchdown ballistically, and mutes the hip torque in stance while readings are stale. Gains came from a (1+λ) evolution search on the blackout plant. **What limits it.** The leg is far heavier than the torso, so the hip cannot place the foot ahead of the body in flight. Each landing therefore happens with the foo…