pendulum-swingup

Inverted pendulum swing-up · solved 13 of 13 agent attempts pass

The task. Classic energy swing-up. The May reference controller fails every seed under this criterion; the June harness found a passing controller in 4 proposals, which is the oracle.

Gate. |wrapped theta| <= 0.15 rad and |theta_dot| <= 0.5 for 25 consecutive steps within 500 steps; success rate >= 0.9 on 40 verifier-only seeds

Axes: dynamics:pendulum, actuation:torque-limited-swingup, n_agents:1

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), 2 min) · base

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

criterion	graded	needed
base:success	1.000	≥ 0.900

How the agent got there: 2 min · 16 tool calls · 3 edit-tool edits · 3 runs of its own Python (tests, experiments, rewrites) · no tuner runs · official verifier 4× (success rates in order: 0.00, 1.00, ?, 1.00).

The agent's final message

Implemented `/app/work/controller.py` with: - Initial torque pulse - Energy-shaping swing-up - Saturated PD upright stabilization Validation: - Visible seeds 0–4: **5/5 success** - Stress test seeds 0–999: **1000/1000 success** - Exact downward equilibrium: successful - Syntax check passedStress test completed successfully: **1,000/1,000 seeds passed**. No follow-up is needed; the final controller remains in `/app/work/controller.py`.

Controller family: energy-based swing-up + LQR catch

Applies a 15-step max-torque pulse, then pumps energy toward the upright energy with torque proportional to energy error times angular velocity, and switches to a PD law once near upright and slow.

Shown: public seed 0 (base); this episode passes the default rung. This episode: final_upright_error 7.31e-06, final_abs_velocity 1.95e-05, success_hold_steps 25, total_reward 11.7.

controller.py · trial pendulum-swingup__wxi82tZ