Skip to content

Autonomous research system for measurable, computer-executable research.

Python
2.1 KB
Raw Blame

Rocket Booster Recovery (Rust) 7000 kg first-contact classical landing research

Improve an auditable deterministic classical controller against the frozen, pure-Rust CPU implementation of the Swordfish C05 6DoF plant. Initial main propellant is fixed at 7,000 kg. The primary metric is landing_success_rate at first landing-leg contact. Success requires touchdown within 5 m, controlled first-contact vertical and leg-tip impact speeds, the established lateral-speed, attitude, and angular-rate safety bounds, and strictly more than 2% main propellant remaining.

Every stratified success rate reuses the same joint predicate. Selection remains multi-objective: overall and hard-OOD coverage, first-contact vertical risk, 2% fuel-gate pass rate, grid-fin saturation, and roll-disturbance rejection remain separate axes. The incubator admits Pareto and new-high candidates from complete, protocol-clean evidence.

Scoring stops at first leg contact. Post-contact springs, damping, rebound, and center-of-mass landing state receive no success credit. A strategy cannot deliberately strike the landing gear at high speed and rely on damping to earn a successful evaluation.

Researchers may modify only controller.rs, controller_config.json, variant.json, and candidate-local .rs modules. The plant, RK4 integrator, contact model, data, evaluator, task harness, Cargo manifests, and Praxist core are frozen. Candidates must pass source-tree static audit, content hashing, and isolated compilation.

CPU Execution Contract

  • Every formal evaluation launches through the Praxist central scheduler's cpu_evaluation profile.
  • Each evaluator defaults to 16 Rayon trajectory workers; one peer may run at most one evaluator concurrently.
  • Server concurrency starts at 8, producing at most 128 planned rollout workers; the central scheduler reduces concurrency under CPU, memory, or I/O pressure.
  • The evaluator and simulator have no GPU backend. Candidates may not add external dependencies, network access, or subprocesses.
  • A candidate's first isolated build may incur one-time Cargo overhead; evidence records scoring wall time separately from build state.