v0.2.1
Major features and improvements:
- Addition of velocity_control actions and support for Gaussian action (#655)
House Assistant Benchmarks Task Changes:
- Collision force measurements fixes. (#684)
- Contact measuring API changes with properly return contacts differentiating between object-on-scene, robot-on-scene, and robot-on-object types of contacts.
- Starting state and the robot arm proper location fix.
- Snap to grasp manager implemented and the grasp manager takes care of the collision groups.
- Refactored observations_to_image to allow variable-sized sensor images. The different image sizes are now tiled side by side. Image and text flip fix. (#687)
- Added actions for setting absolute kinematic state and target joint states. (#685)
- Added separate IK config for the interactive play script.
Refactored measurements code to break down the reward signal and reuse measurements for the reward and success signal.
Habitat Baselines Changes:
- Support rendering images from observation with unequal dimensions.
- Added benchmark_render.py, a new version of benchmark.py which supports rendering videos and saving trajectories to a file.
- Added Gym wrapper with config for which parts of the observation to get.
- Added Sense Plan Act pipeline from Habitat 2.0 paper.
Other improvements and bug fixes:
- Add detail of headless vs. not to the readme (#679)
- Fix to use resume config when training (#695)
- Fix bug with size of the joint observation space.
- Fixed the base transform was not being properly calculated.
- Added interactive play test utility script. This is an important way to test the functionality. (#684)
- Added collision specific measurements.
- Refactored the rearrange simulator class to have less responsibility.
- Record trajectory from the interactive play script for automated testing.
- Added episode generation script for rearrangement tasks under examples folder.
- Fixed properly load in scene config for new ReplicaCAD format.
Corresponding Habitat-Sim v0.2.1 release.
Kudos to our community contributors: @naokiyokoyama.