Repository navigation
2.12.1 (2026-11-12)
This release casts all 2D lidars of a step in one batched pass, lets ORCA groups avoid the obstacles around them, and keeps lidar and collision checks in step with objects moved by hand.
Measured headless on one CPU core of a Linux machine against 2.12.0 (median step time, mean of two runs):
| Time cost | 2.12.0 | 2.12.1 | Speedup |
|---|---|---|---|
env.step(), 1000 robots with 64-beam lidar (usage/27thousand_robot_world) |
2211 ms | 143.9 ms | 15.4x |
env.step(), 1000 robots with 64-beam lidar, ir-sim[fast] (numba) |
2211 ms | 82.9 ms | 26.7x |
env.step(), 1000 dash robots |
79.4 ms | 24.7 ms | 3.2x |
env.step(), 1000 robots with explicit-ID actions |
50.1 ms | 21.6 ms | 2.3x |
Features and Performance
- Batched 2D lidar is up to 26.7× faster. All
lidar2dsensors are cast together in one pass per step, and each scan matches a single sensor's scan bit for bit.pip install ir-sim[fast]installs numba, which compiles the lidar cast for extra speed. Seeusage/27thousand__robot_worldfor details. (#392) - ORCA groups avoid walls, grid maps and moving obstacles. Static objects around the group, including occupied grid-map cells, enter ORCA as obstacles and moving ones as agents, so members no longer pass through them. See
usage/19orca_worldfor details. (#390) - Dash behavior is 3.2× faster for 1000 robots. Dash robots no longer collect the list of all other objects every step, which they never used. Trajectories are unchanged. (#389, based on #386) (@KataDavidXD)
- Explicit-ID actions are 2.3× faster for 1000 robots.
env.step(action, action_id=...)finds each object through a lookup table instead of searching the object list once per ID. Trajectories are unchanged. (#389, based on #386) (@KataDavidXD)
Fix
- Lidar and collision checks immediately see objects moved between environment steps. Before, objects moved with
set_state()orrobot.step()were still searched at their old places until the nextenv.step(); now the geometry tree is rebuilt at the next query. (#393) - GUI backend selection respects user settings and installed toolkits. IR-SIM no longer forces a backend:
MPLBACKENDandmatplotlib.use()are preserved, and Matplotlib can discover Qt6 and GTK backends as well as Tk. (#388) - Image-map tests clean up temporary PNG files on Windows. Four tests no longer fail with
PermissionErrorafter their assertions pass. (#383) (@dhyey-gosa) - Smaller fixes. Missing GUI toolkits produce an installation hint, and lidar range calculations no longer depend on BLAS matrix-size rounding. (#388, #392)
Refactor
- Group behaviors share
target_rolesfiltering. ORCA and SFM groups now default to'obstacle'; settarget_roles: allto include robots outside the group. (#390)
New Contributors
- @dhyey-gosa made their first contribution in #383
- @KataDavidXD made their first contribution in #389 #386