Upkie 3.2.0: Sampling robot states

[stephane-caron]

This release introduces a RobotState type in Python that represents both a state (e.g. the initial state of a simulation) and a distribution around a state we can sample from (e.g. when training a policy by reinforcement learning).

We have also added simulation binaries to each release, which the simulation script will now try to download first, falling back to compilation from source if needed.

Oh, and there is now a template repository to start new agents faster. 🏃 If you are looking for the Pink balancer, it is now an instance of this template, and has moved there.

Added

• MPC balancer: add classes and functions to the documentation
• PID balancer: add classes and functions to the documentation
• PPO balancer: add classes and functions to the documentation
• Robot state class to describe initial state distributions

Changed

• CICD: update release GH action to package simulation spines on various architectures
• Code style: ignore E266 as Doxygen uses ## to document class attributes
• Dependencies: Bump Vulp to 2.1.0
• envs: Default velocity in UpkieServos is now zero
• envs: Observation and action values are float rather than np.float32
• envs: Specify maximum torques in UpkieGroundVelocity
• Simulation script downloads a binary if available, o/w compiles from source (thanks to @pgraverdy)
• whoopsies: Rename try_pid_balancer.sh to start_pid_balancer.sh 😊

Fixed

• envs: typo in UpkieServos dictionary key
• MPC balancer: add missing dependencies to requirements.txt
• PPO balancer: add missing initial state randomization to --training mode

Removed

• Bazel: dependencies based on pip_parse from rules_python
• Pink balancer: moved to a separate repository

---

This discussion was created from the release v3.2: Sampling robot states.