Jiminy-py

This release introduces several experimental reinforcement learning features, along with a new inverse dynamics helper. A few bugs long-lasting bugs and inconsistencies in core have finally been addressed.

New features

* [python/dynamics] Add inverse dynamics helper.
* [gym/toolbox/rllib] Add policy symmetry regularization to PPO.
* [gym/toolbox/rllib] Add discrete hierarchical task space curriculum sampling.

Improvements

* [core] Expose 'Model.computeConstraint' in Python bindings.
* [core] Do not update forward kinematics (position) in 'Robot.computeConstraints'.
* [python/dynamics] Simplify helpers to take advantage of new pinocchio features.
* [gym/toolbox/rllib] Automatically connect to existing cluster if any.
* [gym/toolbox/rllib] Add helper to restore policy without trainer nor ray redis.
* [gym/toolbox/rllib] Add memory tracing for memory leaks in debug mode.

Patches and bug fixes

* [core] Fix memory leaks in Eigen->Numpy Python bindings.
* [core] Rename 'Model.hasConstraint' in 'Model.hasContraints' for consistency.
* [core] Fix motor armatures not properly propagated after setting motor options.
* [python/dynamics] Replace pinocchio methods by jiminy core overloads when appropriate.

Miscellaneous

* [python] Add some missing documentations.
* [misc] Fix python executable detection on manylinux sometimes finding pypy instead of cython.
* [misc] Updating gitignore to support use of virtual environment.

This release introduces a set of reinforcement learning tools based on [ray[rllib]](https://docs.ray.io/en/master/rllib.html). In addition, it is now easier to replay and compare trajectories in different formats (sequence of states, log files, log data, and current simulation).

New features

* [python/simulator] Replay extra logs/trajectories with current simulation.
* [gym/toolbox/rllib] Add dedicated rllib toolbox.
* [gym/toolbox/rllib] Provide [PPO CAPS](https://arxiv.org/pdf/2012.06644.pdf) implementation.

Improvements

* [python/viewer] Refactor Panda3d backend screen refresh and Qt widget to avoid edge cases.
* [gym/toolbox/rllib] Do not provide default log dir since it is error prone

Minor patch to fix viewer issues introduced in previous release.

The interactive mode of `gym_jiminy`, allowing to play with a robot during a simulation using mouse and keyword input, has been greatly improvement in terms of features, reliability, and computational efficiency. Besides, every minor framerate drops of the viewer using Panda3d has been addressed to make the experience more enjoyable. It is now possible to replay more than 10 log files using integrated graphics at solid 30fps.

New features

* [python/viewer] Add option to not lock camera pose but only look-at point when travelling is enabled using Panda3d.

Improvements

* [python/viewer] Display freeflyer external forces by default.
* [python/viewer] Add getter for camera pose.
* [python/viewer] Improve viewer refresh efficientcy by only forcing display update whenever necessary.
* [python/viewer] Add support of frame index to define relative lookup and camera pose.
* [python/viewer] Add repeat capability to 'jiminy_replay' entrypoint.
* [gym/common/envs] More versatile interactive loop helper.
* [gym/common/envs] Add support of gym.wrappers to 'play_interactive'.
* [gym/common/envs] Add 'pause' key to play interactive mode.

Patches and bug fixes

* [python/robot] Fix external forces not properly shared with viewer.
* [python/viewer] Fix interactive loop CPU throttle by adding short sleep in deamon reading stdin.
* [python/viewer] Mix busy loop and timer sleep in precise sleep to release the GIL.
* [gym/common/envs] Fix simulation not stopped when playing ends in play interactive mode.

Miscellaneous

* [gym/examples/rllib] Provide helper to wrap policy as callable[[observation, reward], action].

This release only consists of minor quality-of-life improvements. In addition, a few bugs in Jiminy viewer have been fixed, and a more robust initial guess is provided to the 'impulse' contact solver to improve continuity of contact forces.

Improvements

* [core] Provide robust initial guess to 'impulse' contact solver.
* [gym/common] 'sample' now supports providing both 'shape' and vector 'low'/'high'/'mean'/'scale' if broadcastable.

Patches and bug fixes

* [python/viewer] Fix clicking on node for panda3d Qt widget.
* [python/viewer] Fix seldom lock issues when displaying many robots simultaneously.

Miscellaneous

* [gym/examples/rllib] Start ray backend in local mode in debug.
* [gym/examples/rllib] More robust observation flattening.
* [gym/examples/rllib] Interrupt learning gracefully in case of ray task exception.
* [gym/examples/rllib] Do not replay if recording is disable during training.
* [gym/examples/rllib] Add option to use custom logger.
* [gym/examples/rllib] Add option to save checkpoint every given number of training iterations.

Very small release primarily intended to fix a bunch of bugs introduced in the previous ones. In addition, user-specified standard deviation for model parameters is now relative instead of absolute to be more consistent with the literature, and 'EngineMultiRobot.start' method no longer force reset of the simulation, giving more flexibility to the user, for example to run several simulations using the same model.

Improvements

* [core] 'EngineMultiRobot.Start' no longer reset the engine automatically.
* [core] Biases for random model generation are now relative instead of absolute.
* [python/viewer] Log replay now support overwriting mesh package dir.
* [gym/example] Add option to specify log subdirectory name interactively.

Patches and bug fixes

* [core] Fix computation of center of mass positions and velocities.
* [python/viewer] Fix log replay without velocity and external forces data available.

Miscellaneous

* [misc] Fix CI numpy version issues.