This minor release focuses on improving the simulation speed on GPU for non-batched to moderate batch sizes (~4000) while reduce memory usage across the board. Besides, compilation should be up to x3 faster on Windows OS.
Bug Fixes
- Fix several edge-cases related to glTF parsing. (@kevin9327) (#3342, #3343)
- Preserve textures when replying trajectories. (@duburcqa) (#3395)
- Preserve material textures from collision geoms duplicated as visual geoms when parsing MJCF files. (@MohammadHijjawi97) (#3399)
- Fix 'z_up_to_R' failing when up is omitted. (@Milotrince) (#3396)
- Fix compilation failure in performance mode on Apple Metal. (@duburcqa) (#3421)
- Fix SDF boundary visualization. (@MohammadHijjawi97) (#3425)
- Fix Genesis log messages printed twice when user configures root logging. (@Kashu7100) (#3427)
- Fix the scan lines of a full-rotation spherical raycast pattern. (@MohammadHijjawi97) (#3424)
- Fix pose getters not complying with deterministic mode on GPU. (@Milotrince) (#3387)
- Apply scalar cross-axis coupling to IMU readings. (@cyanseek) (#3436)
- Record the readings of multi-output sensors like IMUs. (@MohammadHijjawi97) (#3419)
- Update the camera extrinsics when the camera moves. (@MohammadHijjawi97) (#3418)
- Fix materials properties defined by TemperatureGrid sensors only applied by the first one. (@Milotrince) (#3439)
- Fix normalisation of log-scaled depth images. (@jeetrex17) (#3442)
- Check broad phase pair overflow when hibernation is enabled. (@Milotrince) (#3415)
Miscellaneous
- Speed up raycasting, tactile sensing and SAP contact queries. (@duburcqa) (#3390)
- Speed up the center of mass computation on GPU for long kinematic chains. (@duburcqa) (#3405)
- Speed up multi-contact detection on GPU for scenes with small batch size. (@duburcqa) (#3404)
- Speed up rigid body simulation on GPU for branched kinematic trees. (@duburcqa) (#3397)
- Speed up rigid body simulation on GPU when the host launch overhead dominates. (@duburcqa) (#3388, #3394, #3433)
- Speed up non-batched rigid simulation on GPU. (@duburcqa) (#3438)
- Reduce the GPU memory footprint of rigid simulation. (@duburcqa) (#3403)
- Reduce memory usage when using heightfield terrain. (@Kashu7100) (#3378)
- Stop enforcing torch's global default device and dtype. (@Kashu7100) (#3431)
- Speed up scene build by caching the description. (@Milotrince) (#3451)
- Replace the dumb theme with a raw theme that prints plain, compact log lines. (@duburcqa) (#3452)
- Upgrade to Quadrants 1.3.3. (@hughperkins) (#3407)