rian-jo / rl-mpc-locomotion

Deep RL for MPC control of Quadruped Robot Locomotion

RL MPC Locomotion

This repo aims to provide a fast simulation and RL training framework for a quadruped locomotion task by dynamically predicting the weight parameters of a MPC controller. The control framework is a hierarchical controller composed of a higher-level policy network and a lower-level model predictive controller.

The MPC controller refers to Cheetah Software but written in python, and it completely opens the interface between sensor data and motor commands, so that the controller can be easily ported to any mainstream simulators.

The RL training utilizes the NVIDIA Isaac Gym in parallel using Unitree Robotics Aliengo model, and transferring it from simulation to reality on a real Aliengo robot (sim2real is not included in this codebase).

Frameworks

Dependencies

• PyTorch - 1.8.1
• Isaac Gym - Preview 4
• Python 3.7

Installation

1. Clone this repository

   git clone [email protected]:silvery107/rl-mpc-locomotion.git
   git submodule update --init

   Or use the --recurse option to clone submodules at the same time.

2. Create the conda environment:

   conda env create -f environment.yml

3. Install the python binding of the MPC solver:

   pip install -e .

Quick Start

1. Play the MPC controller on Aliengo:

   python RL_MPC_Locomotion.py --robot=Aliengo

   All supported robot types are Go1, A1 and Aliengo.

   Note that you need to plug in your Xbox-like gamepad to control it, or pass --disable-gamepad. The controller mode is default to Fsm (Finite State Machine), and you can also try Min for the minimum MPC controller without FSM.

   • Gamepad keymap

     Press LB to switch gait types between Trot, Walk and Bound.

     Press RB to switch FSM states between Locomotion and Recovery Stand

2. Train a new policy: Set bridge_MPC_to_RL to True in <MPC_Controller/Parameters.py>

   cd RL_Environment
   python train.py task=Aliengo headless=False

   Press the v key to disable viewer updates, and press again to resume. Set headless=True to train without rendering.

   Tensorboard support is available, run tensorboard --logdir runs.

3. Load a pretrained checkpoint:

   python train.py task=Aliengo checkpoint=runs/Aliengo/nn/Aliengo.pth test=True num_envs=4

   Set test=False to continue training.

4. Run the pretrained weight-policy for MPC controller on Aliengo: Set bridge_MPC_to_RL to False in <MPC_Controller/Parameters.py>

   python RL_MPC_Locomotion.py --robot=Aliengo --mode=Policy --checkpoint=path/to/ckpt

   If no checkpoint is given, it will load the default runs/{robot}/nn/{robot}.pth

Roadmaps

• MPC Controller
• Quadruped,
• RobotRunner ->
  • LegController,
  • StateEstimator,
  • ControlFSM ->
    • FSM State RecoveryStand,
    • FSM State Locomotion ->
      • ConvexMPCLocomotion ->
        • FootSwingTrajectory,
        • Gait,
        • MPC Solver in C

• RL Environment
• Gamepad Reader,
• Simulation Utils,
• Weight Policy,
• Train ->
  • Vectorized Env,
  • Aliengo Env

User Notes

• Setup a Simulation in Isaac Gym
• Install MIT Cheetah Software
• OSQP, qpOASES and CVXOPT Solver Instructions
• Development Logs

Gallery