Skip to content

v2.2.1 — a fix release

Theme: a long CARLA run no longer dies at a world reload, the OpenCDA entry script works on Windows with --apply_ml, and a frozen SUMO evaluation repeats run to run.

Fixed

  • A failed CARLA world reload stops the run. The automatic world reset used to log "Training may continue" and carry on into a dead world, which ended in a time-out or a native Assertion failed: _episode != nullptr. A reload now runs under a client time-out of at least 120 s (it takes 10 to 35 s on a windowed server), and when it still fails the run stops with a WorldResetError through the usual shutdown path.
  • The first build of an OpenDRIVE world or a town runs under a client time-out of at least 90 s. The single retry is gone: it queued behind the still-running first request, so it could not help a build slower than about 30 s.
  • Sensors are released through the handle that listened to them. Episode reset, adapter teardown, the pre-reload cleanup and a failed spawn setup now stop and destroy each sensor before dropping it. The client warning sensor object went out of the scope could not be reproduced on a windowed RTX 5090 in a 60-episode run, so the fix is covered by unit tests of the release paths rather than by a before-and-after run.
  • --apply_ml runs on Windows. opencda.py imports torch before OpenCV, Open3D and matplotlib, which made import torch fail with WinError 127.
  • The OpenCDA safety manager prints a warning once per change of the warning set, not on every tick of a persisting warning (125 lines against 1 on single_2lanefree_carla). Which warnings are detected is unchanged; whether the car on the freeway map is really off the road is still not established.
  • qt.qpa.screen messages are silenced by default (a user-set QT_LOGGING_RULES wins).
  • SUMO road and junction ids no longer depend on PYTHONHASHSEED. They were built from Python's salted string hash(), the planner breaks ties by road id, and the same frozen cell gave different counters under different hash seeds. They now come from a CRC-32 of the edge name. Results measured before this fix on another hash seed are not reproduced bit for bit.

Known issues

  • The monitor's traffic view does not always match the observation and intent a run produces.
  • platoon_joining_2lanefree_carla does not exercise the joining branches: the merging vehicle reaches the main road 80 to 125 m behind the platoon and about 8 km/h slower. The scenario file is the one upstream ships.
  • QThreadStorage: entry N destroyed before end of thread can still print at exit on a multi-monitor Windows setup.
  • The rest of the v2.1.0 known issues (SUMO spawn-time heading, from_scenario on SUMO only) still apply.