pip install torch numpy scipy matplotlib # python >= 3.10# Table II — nominal docking benchmark (classical + optimal + RL)
python -m rpod_rl.benchmark --backend rcs --methods cw_pd lqr mppi shooting sac ppo \
--sac-ckpt checkpoints/docking_sac/best.pt --ppo-ckpt checkpoints/docking_ppo/best.pt
# Train SAC/PPO from scratch (Sec. IV, Appendix B; ~minutes on one consumer GPU)
python -m rpod_rl.train --env rcs --config full --curriculum --num-envs 1024
# Table III — perturbation robustness sweep (differential J2 / drag / magnetic)
python scripts/perturbation_study.py
# Sec. IV-F / Table IV — refueling attitude hold (RQ3) + plant validation
python scripts/validate_refuel.py
python scripts/refuel_rl_study.py
# Figures
python scripts/fig_tradeoff.py
python scripts/fig_vmax_envelope.pyResult JSONs behind every table/figure are in figures/; trained checkpoints in
checkpoints/. Multi-seed protocol: scripts/seed_sweep.sh + scripts/agg_seeds.py.
World-model agents (TD-MPC2, DreamerV3) run upstream code on our 16-D observation /
4-D action adapters; see third_party_configs/ for the exact configs and commits.
Terminal docking of a 100 kg chaser onto a cooperative target in 500 km circular LEO
(Clohessy–Wiltshire relative dynamics, full rigid-body attitude, RK4). The policy
commands [lateral, longitudinal, yaw, wheel] in [-1,1], mixed onto 8 unidirectional
thrusters (2 N) + a reaction wheel (0.2 N·m). Success requires four simultaneous
port-to-port tolerances (2 cm, 5 cm/s, 5°, 1°/s) inside a 30° approach corridor with a
distance-dependent speed envelope v_max(d) = 0.05 + 0.04 d. The refueling phase models
internal propellant transfer in the mated stack: migrating CoM, time-varying inertia,
İω and CoM-offset thrust torques (rpod_rl/dynamics/refuel_stack.py, validated to
machine precision against closed-form references).