Environments and robots

Choose a backend independently of the planner model. Each integration defines its observations, actions, configuration, and success signal.

Simulation

Environment	Setup target	Example task
LIBERO	libero	--libero-type standard --suite libero_10 --task 0 --seed 0
LIBERO-PRO	libero-pro	--libero-type pro --suite libero_10_task --task 0 --seed 0
LIBERO-Plus	libero-plus	select its task suite with --libero-type plus
ManiSkill	maniskill	--task PickCube-v1 --seed 0
MetaWorld	metaworld	--task reach-v3 --seed 0
robosuite	robosuite	--task Lift --seed 0
Genesis	genesis	--task cube_pick --seed 0
RoboCasa	robocasa	--task OpenDrawer --split target --seed 0
RoboTwin	robotwin	--task beat_block_hammer --task-config demo_randomized --seed 0
RoboLab	robolab	--task BananaInBowlTask --seed 0
RoboDojo	robodojo	task name and evaluation layout as --seed
BEHAVIOR-1K	behavior	activity name and scene instance
HumanCLAW	humanclaw	its humanoid task configuration

MuJoCo environments use EGL, ManiSkill and RoboTwin use Vulkan/SAPIEN, and RoboLab, RoboDojo, and BEHAVIOR use Isaac Sim-based environments. Install and check each target in its own environment.

Mode availability and task coverage depend on the backend. For example, HumanCLAW has its own action-unit vocabulary, and RoboDojo’s Isaac Sim 6.1 configuration runs 48 of its 54 tasks. The backend reference records these distinctions.

Physical robots

Robot	Setup target	Configuration
Franka	franka or franka-polymetis	robot YAML, cameras, calibration, operator
Dual Franka	dual-franka	two controllers and synchronized observations
Piper / dual Piper	piper	ROS environment, robot YAML, operator
UR5e	ur5e	controller identity, Robotiq gripper, cameras, calibration

Real-robot runs use the configured workspace and motion limits and record operator verdicts. Follow the adapter’s setup instructions for the specific controller and camera rig before launching an episode.

Load a robot directly

Outside a wizard-created experiment directory, a single extension can be loaded explicitly:

cd ~/RoboAgentLab
source services/.venv-libero/pi-embodied.env
pi -e packages/embodied/src/robots/libero \
  --libero-type standard --suite libero_10 --task 0 --seed 0 \
  --units=true --max-turns 40 --time-limit 900

Use either this explicit loading pattern or an experiment’s .pi/settings.json. Loading both registers the same tools twice.