Robohive

This pre-release contains two stable task families -- `Kitchen` and `Myo` along with `NPG` baselines.

Kitchen Baselines [(full details)](https://github.com/vikashplus/mjrl_dev/tree/redesign/mjrl_dev/agents/v0.1/kitchen/NPG)
<img width="1001" alt="Screen Shot 2022-03-10 at 11 42 07 PM" src="https://user-images.githubusercontent.com/12837145/157803415-3cb17b16-133c-4e1b-8ebe-e13117690a25.png">

Myo Baselines [(full details)](https://github.com/vikashplus/mjrl_dev/tree/redesign/mjrl_dev/agents/v0.1/myo/NPG)
<img width="1050" alt="Screen Shot 2022-03-10 at 11 45 05 PM" src="https://user-images.githubusercontent.com/12837145/157803682-bc629d01-aeb5-4c17-b78c-20c85a9d3b0e.png">

Known Issues
- Some envs are not serializable (fixed in v0.2dev)
- Arms and HMS suite are missing baselines (planned for v0.2)

- Supports all Gym and Gymnasium versions. Defaults to Gymnasium
- [80ad5bf, 69c1303] Upgraded to MuJoCo 3.1.3
- [777edf3] New teleOperation interfaces (Xbox controllers, spacemouse, keyboard, oculus, controllers) dosssman

FEATURE
- [abc3f70] examine_sim.py upgraded to mujoco bindings andrearosasco
- [69c1303] Checks added for FFmpeg
- [676375f, e4414a1] Upgraded env reset to accept optional seed and kwargs
- [fb90d46, ae51c9e] Better depth rendering accuracy
- [89fe9b2, 65bb9d2, 76b3455] moving towards `env.unwrapped` instead of `env.env...`
- [1cb8334] Loggers can use nested groups
- [777edf3] Xbox 360 controller based tele-operation
Myo Suite
- [15bda40] MyoSuite's base class uses forward call from parent class to avoid duplication
- [9f32f15, 3397e80, 3f74d12] Upgraded to include MyoSuite's MyoChallenge's developments
- [3397e80] Adding the arms and legs models.
- [3397e80] Updating envs to be compatible with gymnasium

BUGFIX
- [82c6ab4] Improved robot class exit. Robot class doesn't delete the persistent robot connection anymore. It relises on the user to close the persistent connection (by calling robot.close() on atleast one of the robot instances) before existing the program. Skipping this for sim-robot shouldn't be a problem. For hardware robot it will leave the robot running. This cleans up the bug that was preventing clean termination due to multiple calls to the robot class.
- [e160593] Removing spurious self collisions in fetch model that was preventing full range of motion
- [7abd6c8, e828a2e] Removing torch dependency from testing.
- [229221a, 7abd6c8] Proper Env seeding post unpickling. Better seeded randomization.
- [3b56770] Removing hardcoded enums from the robot base class.
- [7bea218] Improved MyoSuite sync. SED has a different syntax for -i option on mac.
- [4898218] Camera index 0 gets out of range if no camera is modelled. Defaulting to free camera.
- [315606e] Remove flipping of depth/segmentation buffers sriramsk1999

Tests
- [9e41485] adding a bash script that tests across a few versions
- [0069093] check everything for different gym/gymnasium versions

This checkpoint [overrides](https://github.com/vikashplus/mj_envs/blob/5a8cb3944824abe155efe9bcaf110c46c19c5564/mj_envs/envs/relay_kitchen/__init__.py#L161-L164) the default behavior of the environments from a fixed kitchen to one where the kitchen randomly moves around, making the tasks visually more challenging.

- This release introduces ADROIT benchmarks
- Results for paper - [Learning Complex Dexterous Manipulation with Deep Reinforcement Learning and Demonstrations](https://sites.google.com/view/deeprl-dexterous-manipulation) - was generated using this release.
- Follow the [instructions / codebase of this release](https://github.com/vikashplus/robohive/tree/v0.0) to reproduce results.