Legged Robots that Keep on Learning

Official codebase for Legged Robots that Keep on Learning: Fine-Tuning Locomotion Policies in the Real World, which contains code for training a simulated or real A1 quadrupedal robot to imitate various reference motions, pre-trained policies, and example training code for learning the policies.

Project page: https://sites.google.com/berkeley.edu/fine-tuning-locomotion

Getting Started

• Install MPC extension (Optional) python3 setup.py install --user

Install dependencies:

• Install MPI: sudo apt install libopenmpi-dev
• Install requirements: pip3 install -r requirements.txt

Training Policies in Simulation

To train a policy, run the following command:

python3 motion_imitation/run_sac.py \
--mode train \
--motion_file [path to reference motion, e.g., motion_imitation/data/motions/pace.txt] \
--int_save_freq 1000 \
--visualize

• --mode can be either train or test.
• --motion_file specifies the reference motion that the robot is to imitate (not needed for training a reset policy). motion_imitation/data/motions/ contains different reference motion clips.
• --int_save_freq specifies the frequency for saving intermediate policies every n policy steps.
• --visualize enables visualization, and rendering can be disabled by removing the flag.
• --train_reset trains a reset policy, otherwise imitation policies will be trained according to the reference motions passed in.
• adding --use_redq uses REDQ, otherwise vanilla SAC will be used.
• the trained model, videos, and logs will be written to output/.

Evaluating and/or Fine-Tuning Trained Policies

We provide checkpoints for the pre-trained models used in our experiments in motion_imitation/data/policies/.

Evaluating a Policy in Simulation

To evaluate individual policies, run the following command:

python3 motion_imitation/run_sac.py \
--mode test \
--motion_file [path to reference motion, e.g., motion_imitation/data/motions/pace.txt] \
--model_file [path to imitation model checkpoint, e.g., motion_imitation/data/policies/pace.ckpt] \
--num_test_episodes [# episodes to test] \
--use_redq \
--visualize

• --motion_file specifies the reference motion that the robot is to imitate motion_imitation/data/motions/ contains different reference motion clips.
• --model_file specifies specifies the .ckpt file that contains the trained model motion_imitation/data/policies/ contains different pre-trained models.
• --num_test_episodes specifies the number of episodes to run evaluation for
• --visualize enables visualization, and rendering can be disabled by removing the flag.

Autonomous Training using a Pre-Trained Reset Controller

To fine-tune policies autonomously, add a path to a trained reset policy (e.g., motion_imitation/data/policies/reset.ckpt) and a (pre-trained) imitation policy.

python3 motion_imitation/run_sac.py \
--mode train \
--motion_file [path to reference motion] \
--model_file [path to imitation model checkpoint] \
--getup_model_file [path to reset model checkpoint] \
--use_redq \
--int_save_freq 100 \
--num_test_episodes 20 \
--finetune \
--real_robot

• adding --finetune performs fine-tuning, otherwise hyperparameters for pre-training will be used.
• adding --real_robot will run training on the real A1 (see below to install necessary packages for running the real A1). If this is omitted, training will run in simulation.

To run two SAC trainers, one learning to walk forward and one backward, add a reference and checkpoint for another policy and use the multitask flag.

python motion_imitation/run_sac.py \
--mode train \
--motion_file motion_imitation/data/motions/pace.txt \
--backward_motion_file motion_imitation/data/motions/pace_backward.txt \
--model_file [path to forward imitation model checkpoint] \
--backward_model_file [path to backward imitation model checkpoint] \
--getup_model_file [path to reset model checkpoint] \
--use_redq \
--int_save_freq 100 \
--num_test_episodes 20 \
--real_robot \
--finetune \
--multitask

Running MPC on the real A1 robot

Since the SDK from Unitree is implemented in C++, we find the optimal way of robot interfacing to be via C++-python interface using pybind11.

Step 1: Build and Test the robot interface

To start, build the python interface by running the following: bash cd third_party/unitree_legged_sdk mkdir build cd build cmake .. make Then copy the built robot_interface.XXX.so file to the main directory (where you can see this README.md file).

Step 2: Setup correct permissions for non-sudo user

Since the Unitree SDK requires memory locking and high-priority process, which is not usually granted without sudo, add the following lines to /etc/security/limits.conf:

<username> soft memlock unlimited
<username> hard memlock unlimited
<username> soft nice eip
<username> hard nice eip

You may need to reboot the computer for the above changes to get into effect.

Step 3: Test robot interface.

Test the python interfacing by running: 'sudo python3 -m motion_imitation.examples.test_robot_interface'

If the previous steps were completed correctly, the script should finish without throwing any errors.

Note that this code does not do anything on the actual robot.

Running the Whole-body MPC controller

To see the whole-body MPC controller in sim, run: bash python3 -m motion_imitation.examples.whole_body_controller_example

To see the whole-body MPC controller on the real robot, run: bash sudo python3 -m motion_imitation.examples.whole_body_controller_robot_example