learn-seq

This is a sim2real framework that learn a sequence in simulation and transfer to real-robot.

INSTALL

• install mujoco_py manually

    pip install "mujoco-py=<2.1,>=2.0"
  

• clone learn_seq package

    git clone --recurse-submodules https://github.com/nvuong21/learn-seq.git && cd learn-seq
  

• install dependency for rlpyt

  cpu version

    pip install --requirement rlpyt-cpu.txt
  

  OR gpu with cuda 10.2 (not tested)

    pip install --requirement rlpyt-cuda10.txt
  

• install package for developing

    python setup.py develop
  

  NOTE: python setup.py insttall does not work with the module at the moment

Instruction

• Mujoco model of franka robot is stored in mujoco/franka_sim

• Insertion environment model is stored in mujoco/franka_pih

• Experiments are put in a seperate branch exp

• To create a training experiment named $EXP_NAME, first create a directory in the exp directory, then create a config.py file to setup the configurations (primitives, algorithm, ...). Refer to the exp/example_rlpyt for more detail.

• To run the training experiment for the round peg-in-hole insertion environment using cpus only

    python scripts/train_rlpyt.py -n $EXP_NAME -e round
  

  The training progress, trained model, and training info will be stored in the subdirectory of the $EXP_NAME directory.

• To plot the training performance of all trials in an experiment:

    python scripts/eval_rlpyt.py -n $EXP_NAME --plot-only
  

• To run the evaluation of all trials in an experiment for 5 episodes, and render the progress, run

    python scripts/eval_rlpyt.py -n $EXP_NAME -e 5 --render
  

• To run the evaluation for a single trial in an experiment, add -rn $RUN_NAME at the end

    python scripts/eval_rlpyt.py -n $EXP_NAME -e 5 --render -rn $RUN_NAME
  

TODO

• adjust round_hole_generation.py so that the round_pih.xml can include the generated file
• fixed divide by 0 (quaternion)
• change obs_up_limit and obs_low_limit to be defined relative to the goal -> change peg_pos_range and peg_rot_range specification
• add run_ID to argparse