platoon-delay-noise-20-peak

Vehicle platoon · solved 13 of 14 agent attempts pass

The task. Control arm of the design-V A/B: identical plant, body, oracle and system gate as platoon-delay-noise-20-v, without the vehicle_tracking rig in the gate or the instruction.

The difference between the two tasks' agent runs measures what component-first grading changes.

Gate. the task is solved when the success rate over the grader's held-out seeds is `>= 0.9`

Axes: n_vehicles:20, observation:local-delayed-noisy, leader:hard_brake_train, gate:peak_5m, v-model:control-arm

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.

Best attemptSol (cursor-cli (pre-clock), 10 min) · hard_delay_noise_v1

Graded on the held-out seeds: passes every gate on the grader's held-out seeds.

criterion	graded	needed
base:default	1.000	≥ 0.900

How the agent got there: 10 min · 47 tool calls · 7 edit-tool edits · 6 runs of its own Python (tests, experiments, rewrites) · repository tuner 4× · official verifier 13× (success rates in order: 0.00, 0.00, 1.00, 0.00, 0.00, 0.25, 0.80, 0.31, 0.80, 0.74, 1.00, 1.00, 1.00).

The agent's final message

Implemented `/app/work/controller.py` with deterministic feedforward, delay compensation, local PD feedback, and velocity anchoring. Results: - Visible seeds: 5/5 passed - String peak gain max: 0.766 - Spacing RMS max: 0.304 m - Peak spacing error max: 0.766 m - Additional 100-seed stress test: 100% passed - Syntax check passedTask complete. `/app/work/controller.py` is finalized and verified: 5/5 visible seeds and 100/100 stress-test seeds passed. No follow-up needed.

Controller family: other: nominal-trajectory feedforward + delayed PD spacing control (CACC-style)

Each vehicle applies a precomputed nominal feedforward command for the one-second-headway policy and adds a filtered correction from delay-projected spacing error, relative-velocity residual, and nominal speed tracking, clamped to actuator limits.

Shown: public seed 0 (hard_delay_noise_v1); this episode passes the default rung. This episode: collision no, min_gap 15.7, spacing_error_rms 0.302, spacing_error_abs_max 0.642, string_peak_gain 0.642, episode_steps 600, total_reward -59.4.

controller.py · trial platoon-delay-noise-20-peak__KZpnTtu

FailureSol (cursor-cli (pre-clock), 55 min) · hard_delay_noise_v1

Graded on the held-out seeds: misses the main gate narrowly (success rate between 0.5 and the gate).

criterion	graded	needed
base:default	0.850	≥ 0.900

How the agent got there: 55 min · 122 tool calls · 8 edit-tool edits · 40 runs of its own Python (tests, experiments, rewrites) · repository tuner 1× · official verifier 7× (success rates in order: 0.00, 0.00, 0.57, 0.81, 0.83, 1.00, 0.94).

Pre-clock session: the agent had no clock and may have stopped early believing its hour was over; treat a failure here as a lower bound.

The agent's final message

Implemented `/app/work/controller.py`. Verification: - Visible seeds 0–4: **100% success** - Proxy held-out seeds 101–200: **94% success** - No collisions - Peak spacing error: 4.75 m visible - Worst visible string gain: 1.294 - Syntax check passedTask complete. Final controller is saved at `/app/work/controller.py`; visible success was 100% and held-seed proxy success was 94%. No further action needed.

Controller family: string-stable CACC / platoon

Each vehicle replays the known leader braking script shifted by index times a wave delay, adds low-gain spacing, relative-velocity and acceleration feedback, plus a late-time open-loop balancing wave, and clamps to actuator limits.

Shown: public seed 16 (hard_delay_noise_v1); this episode fails the default rung. This episode: collision no, min_gap 13.6, spacing_error_rms 1.39, spacing_error_abs_max 4.41, string_peak_gain 1.41, episode_steps 600, total_reward -1337. It was the first failing one of 17 public seeds tried.

controller.py · trial platoon-delay-noise-20-peak__4iGE8FR