Jiminy-py

This release introduces first Mujoco-like locomotion learning environment in Jiminy [Ant](https://gym.openai.com/envs/Ant-v2/). An example script successfully training it using TD3 algorithm in Rllib is provided. In addition, several bugs have been addressed, mainly affecting video recording.

Improvements

* [jiminy/simulator] Allow chaining rendering methods.
* [gym/common/utils] Add full support of 'gym.spaces.Tuple'.
* [gym/common/envs] Fix 'render' method not call by env before replaying.
* [gym/common/envs] Enable traveling by default if robot has freeflyer.
* [gym/envs] Add mujoco-like Ant learning environment.

Bug fixes

* [core] Fix 'interpolate' utility.
* [jiminy/viewer] Fix traveling.
* [gym/common/envs] Clip returned observation systematically to avoid out-of-bounds errors.

Miscellaneous

* [gym/toolbox] Consistent addon package hierarchy.
* [gym/example] Refactor examples folder architecture to be easier to extend.
* [gym/example] More generic tools for Rllib. Add Ant TD3 Rllib example.
* [misc] Minor rework of acrobot and cartpole toy models.

This release introduces impulse-like contact model, which has became very popular since the last few years. In broad terms, it maximizes the energy dissipation resulting from contact with the ground, in such a way that the position of the contact points is critically damped with predefined natural frequency (corresponding to solving a LCP internally, see Mujoco [documentation](http://www.mujoco.org/book/computation.html) and [paper](https://homes.cs.washington.edu/~todorov/papers/TodorovICRA14.pdf) for reference). Even though it is not a phenomenological model, it features many very interesting properties, among which allowing very large fixed integration timestep (up to 10ms suing Runge Kutta 4), and smoothing ground reaction forces. Those properties are not essential in robotics to design controllers using classical control theory, but it turns out to be very beneficial when it comes to reinforcement learning, presumably because it makes the state-action mapping much more regular, enabling to learn control policies for complex very dynamic locomotion tasks. Learning examples and tutorials are coming !

In addition, the spring-damper contact model has been greatly simplified to use standard coulomb friction model, and a few significant bugs have been addressed.

New features

* [core] Add new impulse-like contact model.

Improvements

* [core] Check if initial configuration is valid at float precision instead of double then normalized to avoid casting issue in Python.
* [core] Change contacts/bounds spring-damper model to clamp the whole force rather than sole velocity-related component.
* [core] Remove stiction from spring-damper model.
* [python/simulator] Do not fill progress bar if aborted.
* [python/simulator/plot] Add parameter to disable plotting flexibilities. Automatically adjust layout.
* [python/viewer] Add option to only replay specific time interval. Enable to specify unique xyz_offset for replay.

Bug fixes

* [core] Fix inertia bias returning improper matrix.
* [gym/common/envs] Enforce deepcopy of observation and information before returning it to avoid any issue with replay buffer.

Miscellaneous

* [core] Add proper constraints holder to ease implementation of impulse contact model.
* [core] Rename Explicit Euler in Euler Explicit for consistency with Runge Kutta steppers.
* [core] Compute FixedFrameConstraint jacobian and drift LOCAL_WORLD_ALIGNED. Add Baumgarte stabilization.

This release brings minor performance improvement. It is mainly intended to improvement overall code quality and consistency with standard naming conventions (essentially Mujoco terminology). A convenience tool to interpolate trajectories in Lie group space has also been added.

New features

* [core|python/viewer] Provide generic interpolation method for configuration evolution over time. Use it in Python viewer.

Improvements

* [core] Compute subtree coms by default.
* [core] Do not compute 'oMf' for 'JOINT' and 'BODY' frames but rather copy data from 'oMi' and 'oMf'.
* [core] Improvement breakpoint management to avoid skipping controller/sensor updates.
* [core] Return static variables for Robot 'getControlLimit' and 'getArmatures' to avoid memory allocations.
* [core] Set 'command' and 'u_custom' to zero automatically before calling callbacks.
* [core] Improve vector 'clamp' utility to avoid memory allocation.
* [core/python] Make sure registered variables vector is 1D.
* [core|python] Add option to monitor motor efforts and plot them by default instead of command.
* [python/robot] Enable setting options only partially.
* [python/dynamics] Add optional flag to update collisions. Fix com jacobian not computed even if requested.
* [python/simulator] Do not forbid to render system at if no simulation is running.
* [gym/common] Ensure float64 action buffer for compatibility with register_variable telemetry.
* [gym/common/envs] Increment 'num_steps' only at the end of successful step, not right before integration.
* [gym/common/envs] Set robot in neutral configuration at init.

Bug fixes

* [core] Fix edge case in controller breakpoints management.
* [gym/common/envs] Fix edge case for robot without sensors.
* [gym/common/envs] Do not enforce position/velocity/effort bounds at _setup to avoid inconsistent observation/action spaces bounds.

Miscellaneous

* [misc] Move to C++17 for dependencies.
* [core] Rename 'rotorInertia' in 'armature' to be consistent with mujoco and clearer that it is joint side.
* [core] Rename 'EffortLimit' in 'ControlLimit' to avoid redundancy with pinocchio::Data and explicit it is motor-side.
* [core] Rename 'ControlLimit' in 'CommandLimit' for consistency.
* [core] Rename 'u_command' in 'command' to clarify it does not have to be an effort.
* [core] Add dedicated 'u_custom' field for controller's internal dynamics.
* [core] Replace 'u' with more specific variable name when possible.
* [core] Split evaluation of custom internal dynamics from position/velocity bounds and flexibility dynamics.

Another small release consisting mainly in bug fixes. In addition, a new type of pre-defined visco-elastic coupling force has been added, acting independently in each direction instead of only along coupling direction.

New features

* [core] Add non-directional visco-elastic coupling force (acting independently in each direction).

Improvements
* [core] Fix adaptive RK dopri stepper MAX_FACTOR wrongly scaled by SAFETY factor.
* [gym/common/envs] Do not log internal steps in debug mode to keep log consistent with experience replay.

Bug fixes

* [core] Fix visco-elastic directional coupling force.
* [core] Fix pinocchio_data updated with joint forces not taking into account the external forces.
* [core] Fix telemetry logging failed internal stepper steps.
* [core] Fix wrong computation of multi-systems coupling forces.
* [python/viewer] Fix replay using rigid instead of flexible model.
* [python/viewer] Fix mesh path on windows.

Miscellaneous

* [gym/common] Rename 'size' argument of 'sample' method in 'shape' for consistency with standard API.

Tiny release to fix a critical bug causing corruption of sensors data (IMU, Force Sensors, and Contact Sensors).

Bug fixes

* [core] Fix joints accel/force restored right after computations instead of before sensor update.

Miscellaneous

* [misc] Check no memory allocation by Eigen during simulation.
* [misc] Update Eigen to latest release (3.3.9) on CI.

Simulation speed has been increased by at least 15%, mainly avoiding memory allocation during simulation, postponing accelerations and forces updates until successful stepper step, and not computing all terms for internal stepper steps.

* [core] Avoid algebraic loop in internal stepper steps by postponing accel/force updates after successful outer step.
* [core] Move centroidal kinematics computations in stepper outer loop for efficiency.
* [core] Assume slow variation of subtree inertia during one integration step for efficiency.
* [core] Avoid memory allocation for 'computeAcceleration', 'updateTelemetry' and 'computeAllEffort'.
* [misc] Fix double_pendulum c++ example using Eigen::Ref for controller handles instead of vectorN_t.