Official Implementation of Agentic Skill Discovery (ASD)
This is the official implementation of agentic skill discovery, a framework driven by LLMs to generate task proposals based on the scene description and the robot's configurations. For each task, reinforcement learning processes, guided by LLM-derived success and reward functions, develop the necessary policies. An independent vision-language model ensures the reliability of learned behaviors.
The simulation is based on Isaac-sim.
- DOING Currently organising code to run inside docker without manual configurations
apt-get update
apt-get install -y tmux zsh wget git python3 python3-pip vim openssh-client
sh -c "$(wget https://raw.githubusercontent.com/ohmyzsh/ohmyzsh/master/tools/install.sh -O -)"
ssh-keygen -t ed25519 -C "xfz.zhao@gmail.com"
eval "$(ssh-agent -s)"
ssh-add ~/.ssh/id_ed25519
cat ~/.ssh/id_ed25519.pub
Copy the last line and add to SSH keys: https://github.com/settings/keys.
cd orbit
export ISAACSIM_PATH=/isaac-sim
export ISAACSIM_PYTHON_EXE="${ISAACSIM_PATH}/python.sh"
export ORBIT_ROOT_DIR=$(pwd)
export ZEROHERO_ROOT_DIR=$ORBIT_ROOT_DIR/zero_hero
export PYTHON_EXE="${ISAACSIM_PATH}/python.sh"
export OPENAI_API_KEY=
/isaac-sim/kit/python/bin/python3 -m pip install "usd-core<24.00,>=21.11"
./orbit.sh --install
./orbit.sh --extra rsl_rl
ln -s ${ISAACSIM_PATH} _isaac_sim
cd zero_hero
git pull
pip3 install -r requirements.txt
...
value = term_cfg.func(self._env, **term_cfg.params) * term_cfg.weight * dt
# Add new line here
value = value.squeeze()
# update total reward
self._reward_buf += value
...
Launch train.py
under zero_hero
directory to examine whether oribt is successfully installed.
../orbit.sh -p rsl_rl/train.py --task Franka_Table --headless --num_envs 4096 --max_iterations 10
Normally, the training starts with a ~50k frames/s on a A100 gpu card.
python3 learn.py task="Reach cube A." num_envs=4096 memory_requirement=32 min_gpu=90 temperature=0.7
Install Everything Inside Isaac Sim Docker
- Pull docker image and run
docker run --name isaac-sim --entrypoint bash -it --gpus all -e "ACCEPT_EULA=Y" --rm --network=host \
-e "PRIVACY_CONSENT=Y" 134.100.39.10:32000/isaacsim:2023.1.1
Reference: https://catalog.ngc.nvidia.com/orgs/nvidia/containers/isaac-sim
- Inside the launched docker container, execute
cd /isaac-sim
./runheadless.native.sh
If normal, no obvious error.
Extra information:
NVIDIA-SMI 550.54.15 Driver Version: 550.54.15 CUDA Version: 12.4
- Create
yaml
with
./mlpod.py --yaml isaacsim2.yaml --user gaede --pod isaacsim --image 134.100.39.10:32000/isaacsim:2023.1.1 --gpumem 60 --env ACCEPT_EULA=Y PRIVACY_CONSENT=Y -- /bin/bash
-
Edit
isaacsim2.yaml
by replacing the twoY
with"Y"
(as string instead of as bool value). -
Create pod with the
yaml
file
alias kubectl k
k create -f isaacsim2.yaml
k attach -it isaacsim2
- Inside the attached pod, run
cd /isaac-sim
./runheadless.native.sh
May run into errors:
2024-04-16 08:45:58 [1,012ms] [Warning] [omni.platforminfo.plugin] failed to open the default display. Can't verify X Server version.
2024-04-16 08:45:58 [1,032ms] [Error] [carb.graphics-vulkan.plugin] VkResult: ERROR_INCOMPATIBLE_DRIVER
2024-04-16 08:45:58 [1,032ms] [Error] [carb.graphics-vulkan.plugin] vkCreateInstance failed. Vulkan 1.1 is not supported, or your driver requires an update.
2024-04-16 08:45:58 [1,032ms] [Error] [gpu.foundation.plugin] carb::graphics::createInstance failed.
2024-04-16 08:45:58 [1,608ms] [Error] [carb.graphics-vulkan.plugin] VkResult: ERROR_INCOMPATIBLE_DRIVER
2024-04-16 08:45:58 [1,608ms] [Error] [carb.graphics-vulkan.plugin] vkCreateInstance failed. Vulkan 1.1 is not supported, or your driver requires an update.
2024-04-16 08:45:58 [1,608ms] [Error] [gpu.foundation.plugin] carb::graphics::createInstance failed.
2024-04-16 08:45:59 [2,155ms] [Error] [omni.gpu_foundation_factory.plugin] Failed to create any GPU devices, including an attempt with compatibility mode.
Extra information:
NVIDIA-SMI 550.54.14 Driver Version: 550.54.14 CUDA Version: 12.4