v3.0.0b19

What's Changed

• [Feature] Upgrade pytorch_kinematics by @StoneT2000 in #854
• Fixed final_info having wrong elapsed_steps value after reset and final_observation being wrong on CUDA by @AlexandreBrown in #828
• Fix agents' qpos type to be floats by @shaido987 in #856
• [Docs] Update docs on IL baselines regarding standard IL datasets and how to replay them, remove old examples.sh files by @StoneT2000 in #861
• YCB, RCAD scene builder updates, demo_manual_control by @arth-shukla in #859
• [BugFix] update the checksum for ycb dataset by @StoneT2000 in #865
• [BugFix] Fix slow CPU sim and fix the CPU sim benchmark code to be faster by @StoneT2000 in #870
• [Feature] Added RGB DP baseline and DrawTriangle/SVG task by @StoneT2000 in #876
• [Feature] Update PPO baselines for pick single YCB and remove wandb-entity CLI arguments by @StoneT2000 in #863
• Fix crash in CPU backend for multi-body Actors by @shaido987 in #880

New Contributors

• @AlexandreBrown made their first contribution in #828
• @shaido987 made their first contribution in #856

Full Changelog: v3.0.0b18...v3.0.0b19

v3.0.0b18

What's Changed

• [BugFix] Fix off by one error in saving GPU replayed trajectories by @StoneT2000 in #842

Full Changelog: v3.0.0b17...v3.0.0b18

v3.0.0b17

What's Changed

• [Feature] Support pd_joint_pos to pd_ee_pose controller action conversion. Fix bug where motion planning demos were using GPU sim backend to replay by default and weren't working. by @StoneT2000 in #834
• [Feature] Handle more flexible observation modes, support users to check if ground truth state data is requested by @StoneT2000 in #835
• [BugFix] Fix initial poses of static objects in AI2THOR scenes to be not zero in CPU and GPU sim. by @StoneT2000 in #836
• [Feature] Remove extra tensor init in PPO code. by @StoneT2000 in #837
• [BugFix] Fix bug where replay trajectory tool would use torch tensor seeds and try to save them in JSON by @StoneT2000 in #840

Full Changelog: v3.0.0b16...v3.0.0b17

v3.0.0b16

What's Changed

• [Docs/BugFix] Fix bug with sac replay buffer incorrectly filling up after max and add new PPO baselines for many tasks by @StoneT2000 in #731
• Fix documentation for torch_deterministic parameter in ppo.py by @songyuc in #745
• [Feature] Support applying external forces on objects by @StoneT2000 in #752
• Adding Action Chunking with Transformers (ACT) to baselines by @ywchoi02 in #640
• [BugFix] action conversion to end-effector target control having shaky behavior by @StoneT2000 in #756
• [BugFix] fix bug with gpu benchmarking script when running maniskill environments due to missing property by @StoneT2000 in #762
• [Feature] Batched and simpler API for setting locked motion axes by @StoneT2000 in #763
• [Feature] support set mass on GPU by @StoneT2000 in #792
• fixing minor typos in docs by @vncntt in #801
• Remove unnecessary conversion of info to NumPy (#791) by @hesic73 in #796
• [BugFix] Fixed dtype error when running TD-MPC2 on anymalc-reach by @t-sekai in #808
• ManiSkill-HAB to docs by @arth-shukla in #759
• [Feature] Cleanup and standardize the imitation learning baselines and datasets by @StoneT2000 in #811
• [Feature] cleaner baseline scripts for ACT baseline and DP by @StoneT2000 in #815
• [BugFix] Improve PPO default parameters and fix bug with lazy init sensors/cameras by @StoneT2000 in #816
• [Docs] update gallery with more videos by @StoneT2000 in #819
• [Feature] Support state+<visual_textures> mode by @StoneT2000 in #821
• Fix typos by @vncntt in #802
• Fix: vector-gymnasium-wrapper -- fixing 'call' method -- getattr(self.env...) … by @imishani in #783
• [Feature] Support XArm6 with robot iq gripper by @StoneT2000 in #820
• [BugFix] Fix usage of the MS_ASSET_DIR variable by @StoneT2000 in #824

New Contributors

• @songyuc made their first contribution in #745
• @ywchoi02 made their first contribution in #640
• @vncntt made their first contribution in #801
• @imishani made their first contribution in #783

Full Changelog: v3.0.0b15...v3.0.0b16

v3.0.0b15

What's Changed

• [BugFix] Fix bug with record episode wrapper for single CPU envs by @StoneT2000 in #723
• [Feature] Substep Rendering by @StoneT2000 in #724

Full Changelog: v3.0.0b14...v3.0.0b15

v3.0.0b14

What's Changed

• [Version] Version Bump to b14 by @StoneT2000 in #721
• [BugFix] Fix bug with missing init py file in panda stick motion planning folder

Full Changelog: v3.0.0b13...v3.0.0b14

v3.0.0b13

v3.0.0b13 Release

Notable changes/additions:

• A new PPO based on LeanRL has been provided under examples/baselines/ppo_fast. PPO training is much much faster now (solve PickCube in a minute! Solve PegInsertionSide in about 1 hour!)
• Scripts for RL based data generation are provided in scripts/data_generation/rl.sh now
• Multiple new tasks have been added including a few bimanual humanoid tasks
• Some minor bugs have been fixed and simulation speed has been improved in a number of ways.
• Much more documentation on domain randomization tools and how to use them for your own environments: https://maniskill.readthedocs.io/en/latest/user_guide/tutorials/domain_randomization.html
• RGB(D) based baseline for SAC added.
• Fix bug with the pd_ee_delta_pose controller not working when using PPO code due to in place modification in that controller.
• Upgraded documentation showing more information now about all tasks available by default in ManiSkill. Task documentation now automatically details whether the task has demos, dense/sparse rewards, as well as its episode length
• Many more demonstrations have been added with more coming, demonstrations are now all placed in the same hugging face datasets repo here: https://huggingface.co/datasets/haosulab/ManiSkill_Demonstrations.

UnitreeG1TransportBox-v1_rt.1.mp4

Full Changelog:

What's Changed

• [Fix] Use same indices for root vel setter as base uses for articulations by @arth-shukla in #665
• [BugFix] Fix bug with digital twin rgb overlay due to new to tensor system and display availabe obs modes if chosen one is not correct by @StoneT2000 in #668
• [BugFix] Fix record episode wrapper storing all parallel env episode seeds instead of just one. by @StoneT2000 in #671
• [Bug] Fix: (unideal) only do .step() if not doing partial reset by @arth-shukla in #672
• [BugFix] Fix demo_vis_segmentation script by @StoneT2000 in #675
• [Feature] Unified states for heterogenous simulated envs by @StoneT2000 in #676
• [Docs] Add docs on simulation properties like object textures and controller properties that can be dynamically updated during runtime. by @StoneT2000 in #681
• [Feature] Support normals and albedo in observation data by @StoneT2000 in #682
• [Feature] Added batched input support for look_at matrix construction by @Xander-Hinrichsen in #678
• [Feature[ Update interactive_panda.py to allow using different viewer shaders by @StoneT2000 in https://github.com//pull/687
• [Feature] Extremely fast PPO by @StoneT2000 in #669
• [BugFix] Pose creation always creates a new object now and does not duplicate tensors. by @StoneT2000 in #688
• [BugFix] Fix PD EE Delta Pose controller that would sometimes do in place modification of user given action by @StoneT2000 in #689
• [BugFix] Update ycb actor builder to include every object in YCB dataset, not just the ones that can be picked up. by @StoneT2000 in #690
• [BugFix] Fix bug where consistency check for state in env state recording assumes actors/articulations exist by @StoneT2000 in #691
• [Feature] Organized data generation scripts by @StoneT2000 in #692
• [Feature] SAC RGB(D) Baseline and improved memory efficiency by @StoneT2000 in #698
• Fix flatten_state_dict with empty dictionary by @younik in #700
• [BugFix] Fix issue with demo_manual_control.py by @StoneT2000 in #702
• [Feature] Auto-generated task documentation by @StoneT2000 in #703
• Added custom task (PullCubeTool-v1) with motion planner + Motion Planner for LiftPegUpright and PullCube by @Viswesh-N in #641
• [Docs] Update docs on PullCubeTool-v1, and upload new demonstrations to HF for motion planning generated datasets by @StoneT2000 in #704
• [BugFix] Fix kinematics that sometimes will skip an active joint by @StoneT2000 in #705
• [BugFix] fix bug when a torch based pose instead of sapien pose is passed to RRT planning code by @StoneT2000 in #706
• [BugFix] Remove extra warnings and clarify why initial poses should be set by @StoneT2000 in #709
• [Feature] More humanoid/bimanual manipulation tasks, fix G1 robot's frictions, optimize robot collisions for franka and G1. Add G1 with head camera, allow record with info for parallel envs by @StoneT2000 in #715
• [BugFix] Fix bug for recording episodes with CPU envs by @StoneT2000 in #718
• [Docs] Documentation on even more Domain Randomization by @StoneT2000 in #719

New Contributors

• @younik made their first contribution in #700
• @Viswesh-N made their first contribution in #641

Full Changelog: v3.0.0b12...v3.0.0b13

v3.0.0b12

v3.0.0b12 Release

Big changes include

• RoboCasa scenes integrated into ManiSkill, simulating easily 500-1000 RGBD FPS. Run python -m mani_skill.examples.demo_random_action -e "RoboCasaKitchen-v1" --render-mode="human" --shader rt-fast to try it!
• Batched RNG is now recommended to be used for scene loading in tasks to ensure CPU/GPU sims can generate the same exact objects/geometries/textures for a task when given the same list of seeds. Read https://maniskill.readthedocs.io/en/latest/user_guide/concepts/rng.html to learn more about the details
• Motion planning solutions have CPU parallelization options now
• Fix bug with point cloud data being scaled down twice
• Add a warning if you don't add initial poses to builders. This is strongly recommended to do now by overriding def _load_agent to initialize a robot pose and to set initial poses of other objects such that they don't intersect if they were to spawn at the initial poses. This is absolutely necessary for fast and accuracte GPU simulation, and fixes a number of issues with the open cabinet task.
• Major bug fixes with the RNG that were introduced in 3.0.0b11. Namely envs wouldn't randomize at all despite resets sometimes. Please upgrade to 3.0.0b12 instead of using b11.

What's Changed

• docs: update README.md by @eltociear in #594
• add missing 'sensor_data' handling in parse_visual_obs_mode_to_struct by @hesic73 in #606
• Update docs by @FarukhS52 in #614
• [BugFix] ControllerDict prevented before_simulation_step on child controllers by @RoboNuke in #613
• [Docs] Update docs on demonstrations and imitation learning by @StoneT2000 in #618
• [Docs] Fix Typos in docs by @VaibhavWakde52 in #622
• [BugFix] Agent registration not working when overriding a robot uid by @rechim25 in #623
• [Feature]: Add CPU multiprocessing for Panda Motionplanner's Trajecto… by @chenyenru in #611
• [Feature] Habitat Rearrange tasks + Bug fixes use due to bad default initial poses by @arth-shukla in #529
• [BugFix] Add error handling for Panda's Motionplanner by @chenyenru in #627
• [Feature] Reproducible batched RNG support, fix state representations of various tasks by @StoneT2000 in #619
• [BugFix] Fix OpenCabinetDrawer-v1 by @StoneT2000 in #630
• [BugFix] Fix bug in SAC code and remove some old code in RL examples by @StoneT2000 in #639
• [Feature] Support the Robocasa scene dataset by @StoneT2000 in #648
• [(Small) Feature]: small build util changes, extra humanoid keyframes + attrs by @arth-shukla in #651
• [BugFix] Improve GPU sim speed and new docs on initial pose handling by @StoneT2000 in #655
• [BugFix]: improve perf for batched_rng by @arth-shukla in #656
• [BugFix] Fix bug where sometimes the same episode rng is given if reconfiguration freq is 1 but enhanced determinism is false and no seed is given by @StoneT2000 in #659

New Contributors

• @eltociear made their first contribution in #594
• @hesic73 made their first contribution in #606
• @FarukhS52 made their first contribution in #614
• @RoboNuke made their first contribution in #613
• @VaibhavWakde52 made their first contribution in #622
• @rechim25 made their first contribution in #623

Full Changelog: v3.0.0b10...v3.0.0b12

v3.0.0b10

v3.0.0b10 is released

This new version introduces a lot of new features

• Ant walk/run control tasks added and solvable via visual/state based PPO
• New, clean, systematic wrappers/design for fairly evaluating algorithms in simulation for RL, IL etc.
• All end-effector / IK based controllers are now GPU parallelized. Big thank you to @LemonPi for help on batched IK control via their pytorch kinematics library
• four Real2sim environments from SIMPLER have been added, cleaned up, and show good correlation between sim and real success rates and low MMRV relative to original CPU implementations. Evaluation is now 60-100x faster than the real world!
• Trajectory conversion tool for the panda arm is now much more accurate. Tasks like peg insertion / plug charger can convert from pd joint position to any other action space with a much higher success rate.
• Panda wristcam robot had an incorrect initialization on the table scene which has been fixed.
• A base drawing simulation environment with no tasks or demos at the moment
• TwoRobotPickCube-v1 has a new more optimized reward function
• Updated baselines to adhere to new standards of evaluation metrics and reporting
• New sim+rendering benchmarks
• Fix various bugs

Demos of the new drawing environment and real2sim environments below

drawing_special_s_symbol_task.mp4

real2sim_spoon_task.mp4

What's Changed

• [BugFix] Fix bug with processing segmentation images for older torch versions by @StoneT2000 in #532
• [BugFix] Bump sapien dependency in tdmpc2 baseline for mani_skill update by @chenyenru in #527
• Added MS Ant Walk and Run Tasks by @Xander-Hinrichsen in #533
• [Feature] Updated TD-MPC2 evaluation and fixed some bugs by @t-sekai in #538
• Added behavior cloning baselines by @arnavg115 in #531
• [Feature] Align and simplify policy/ML evaluation in ManiSkill by @StoneT2000 in #544
• [Feature] Massively improve success rate of pd joint pos / delta pos conversion to pd ee delta pos/pose controller by @StoneT2000 in #549
• [Feature] Clean up trajectory replay code and more docs on the tool by @StoneT2000 in #552
• panda wristcam initialization added by @jamesahou in #553
• [Feature] Drawing Simulation Base Environment by @StoneT2000 in #555
• [BugFix] Fix CPU articulation net contact forces being wrong shape by @StoneT2000 in #556
• [Feature] Real2Sim Eval Digital Twins by @StoneT2000 in #536
• Solved PPO for TwoRobotPickCube-v1 by @jamesahou in #562
• Behavior cloning bugfixes by @arnavg115 in #565
• [Docs] New benchmarking scripts by @StoneT2000 in #568
• Benchmarking code 2.0 by @StoneT2000 in #573
• [BugFix] Fix panda_v2_gripper.urdf visual mesh suffix (dae->glb) by @kh11kim in #575
• [Docs] More rendering benchmarking by @StoneT2000 in #580
• [Docs] Sim+rendering Benchmarking, matching visuals by @StoneT2000 in #584
• [Docs] Paper upload and update documentation and READMEs by @StoneT2000 in #585
• [Docs] fix typos in paper header by @StoneT2000 in #586
• [Docs] Update docs on controllers and fix video link by @StoneT2000 in #589
• [Feature] Updated TD-MPC2 with CPU mode and new shared metrics by @t-sekai in #578
• [Version] v3.0.0b10 by @StoneT2000 in #591

New Contributors

• @chenyenru made their first contribution in #527
• @jamesahou made their first contribution in #553
• @kh11kim made their first contribution in #575

Full Changelog: v3.0.0b9...v3.0.0b10

v3.0.0b9

What's Changed

• [BugFix] Fix render all mode by @StoneT2000 in #521

Full Changelog: v3.0.0b8...v3.0.0b9