Week 13 — Phase 4 Complete, Showcase Prep Begins 7 Days Early

Forest navigation now works cleanly — 0 body penetrations in 700-step play with 8 drones weaving through 40cm tree trunks. Phase 4 is complete and Phase 5 (Showcase Prep) starts 7 days early.

Milestones

• Phase 4 COMPLETE. All seven M3 criteria pass (with the corrected body-radius obstacle metric). Production checkpoint locked in at logs/skrl/ggswarm/p4/2026-04-06_21-09-24_ppo_torch/.
• Clean forest run. p4-revert-4 checkpoint (reward 66.83 / ep_len 307.74) trains slightly better than the Mar 31 baseline. 0 body penetrations, +0.037m min body clearance, 0 stuck drones — vs 128 penetrations and piled-up drones at the start of the week.
• Reverted the learned-obstacle-avoidance experiment. Six commits of obstacle penalties, obs columns, and curriculum gave zero measurable benefit across six GCE runs. Archived to experimental/learned-obstacle-avoidance; main is back on goal deflection.
• Boids-style flock alignment for deflection. Replaced pure radial deflection with a 70/30 lateral/radial blend, side picked from mean K-nearest neighbor velocity. Drones now coordinate which side of a cylinder to dodge.
• Forest cylinders widened to 40cm diameter for visual realism (young tree trunks); bumped cbf_obstacle_d_safe to 0.60m to compensate.

Forest Play Results (p4-revert-4, 700 steps × 8 drones)

Key Bugs Squashed

• Cfg drift regression. Two retrains collapsed to reward 24/7 vs the 63 baseline. Env code was character-identical to Mar 31 — culprit was 100% drift in cbf_d_safe, cbf_max_correction, collision_radius, and dropout_enabled. A bad recovery attempt also reproduced the p3-16 unclamped-CBF flip by raising cbf_max_correction to 0.50 (true Mar 31 value was hardcoded _MAX_CORRECTION = 0.15 in cbf.py).
• Deflection used goal position, not drone position — drones whose slots landed just outside the deflection radius drifted into cylinders under formation pressure and were never protected.
• Base goal advanced unconditionally — stuck drones had their goal run away up to 4.12m ahead, drowning out lateral deflection. Fix: cap with new forest_max_goal_lead = 0.5m.
• Historical “0 hits” was a measurement bug. p4-forest-14/15/16 and the M3 gate never included the 0.10m drone radius. Re-measured every historical forest run — all grazed cylinders by ~5cm. Re-stated in phase4_stress_testing.md § 4 and § 5.5.

Next Steps

• Phase 5A: add --tron flag to play.py that calls setup_tron_environment() after gym.make(). ~30 lines, enables the first Tron-styled forest video same session.
• HD demo capture with the trees2 forest config and the production checkpoint.
• Phases 5B–D: Tron visual debug, formation morphing scene sequencer, full cinematic trailer (~2:30 at 1080p 60fps).
• Post-capstone: downwash penalty for vertical-stacking at spawn (real-world fidelity, not on critical path).

Challenges

• A week of debugging that, on paper, produced one reverted experiment and one cfg fix. The real lesson: hardcoded constants in module files (_MAX_CORRECTION in cbf.py) silently became part of the baseline and weren’t visible in cfg diffs. Promoted them to function parameters per the project’s “no magic numbers” rule.
• 17 days remain to the Apr 24 deadline — Phase 5 starting 7 days early gives real headroom for the cinematic showcase work.