Skip to content

CHANGELOG.md

Latest commit

 

History

History
62 lines (51 loc) · 3.29 KB

CHANGELOG.md

File metadata and controls

62 lines (51 loc) · 3.29 KB
Raw

Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog and this project adheres to Semantic Versioning.

Unreleased

Added

  • Eight drones on the 8x8 obstacle course, solved: 10 of 10 seeds bring every drone home and hold, in 21 steps against a best-found central plan of 17. The fix was exploration, entropy_coef: 0.06 on that profile alone, after feasibility, congestion, geometry, longer training, curriculum transfer and epoch count were each measured and ruled out; the negative results live in configs/fly-fleet8.yaml.

  • Arrival now means arrived and stayed. A drone that leaves a goal it reached is disqualified for the episode, shown amber in the viewer, and counted separately as occupying; end-of-run labels, live readout, and board colors all answer the same question.

  • scripts/plan_optimum.py: space-time A* over prioritised orderings, the feasibility bound the learned policies are measured against.

  • Browser viewer rebuilt as an industrial yard flown by a real drone model (VR Drone by Dave404, CC-BY, with a procedural quadcopter as fallback): runtime canvas textures, sun shadows, occlusion fade kept, helipads with owner rings, X, Y, Z and Default camera flights, eased continuous zoom. The dusk-city variant is preserved on the viewer-city-b branch.

  • README landing: recorded eight-drone rollout as a looping inline animation, plus stills shot from the shipped build.

  • Opt-in training knobs, off by default and measured before shelving: running return normalisation with a forgetting horizon, entropy annealing.

  • Multiple drones sharing one grid, sized by num_drones, with shared policy weights.

  • Path finding: vertex, swap and stationary conflicts detected and refused each step.

  • Obstacles drawn from obstacle_density, with four local sensor flags per drone.

  • Safety Controller with geofence and separation rules, the only component permitted to veto.

  • Training configuration via --config, applied to the agent rather than discarded.

  • configs/env-prod.yaml as a loadable production profile.

Changed

  • /predict takes one observation row per drone and returns action indices and names.
  • Environment migrated from the deprecated gym package to gymnasium.
  • Observation space bounds are per-dimension.
  • Imports normalised to a single package layout, which unbroke the production image.

Fixed

  • CI on all three workflows: the gymnasium import, an httpx incompatibility, and the missing editable install in the Docker image.
  • observation_space declared one scalar bound across every dimension, so every step reported a spurious drift error.
  • /predict accepted no valid input: the schema wanted a mapping, the policy wanted numbers.
  • tests/conftest.py swallowed import errors and substituted stubs, masking both bugs above.
  • training_reward was declared but never observed, leaving its Grafana panel empty.

0.1.0 - 2025-09-08

Added

  • Initial production baseline, observability, deployment configs and utilities.