LW-BenchHub Robot Overview

Robot Configuration Summary Table

Robot Name	--robot Parameter	Function Summary	Source/Type
G1-Controller	`G1-Controller`	Basic control interface for Unitree G1 hardware	Unitree G1 Humanoid Robot
G1-Loco-Controller	`G1-Loco-Controller`	Controller with advanced locomotion module	Unitree G1 + Locomotion Module
G1-Hand	`G1-Hand`	Humanoid robot with dual-arm manipulation and base mobility	Unitree G1 Humanoid Robot
G1-Loco-Hand	`G1-Loco-Hand`	Humanoid robot with enhanced locomotion and leg control	Unitree G1 + Locomotion Control
G1-RL	`G1-RL`	RL-optimized humanoid with single-arm manipulation	Unitree G1 for Reinforcement Learning
G1-FullHand	`G1-FullHand`	Humanoid robot with full dexterous hands (all fingers DOF)	Unitree G1 with Full-Fingered Hands
PandaOmron-Rel	`PandaOmron-Rel`	Mobile manipulator with relative position control	Franka Panda + Omron Mobile Base
PandaOmron-Abs	`PandaOmron-Abs`	Mobile manipulator with absolute position control	Franka Panda + Omron Mobile Base
DoublePanda-Rel	`DoublePanda-Rel`	Dual-arm system with relative position control	Dual Franka Panda Setup
DoublePanda-Abs	`DoublePanda-Abs`	Dual-arm system with absolute position control	Dual Franka Panda Setup
LeRobot-RL	`LeRobot-RL`	RL-optimized lightweight collaborative arm	LeRobot SO100/101 Arm
AbsJointGripper-RL	`AbsJointGripper-RL`	RL-optimized absolute joint gripper variant	LeRobot SO100/101 Arm (absolute, RL)
BiARM-RL	`BiARM-RL`	RL-optimized dual-arm collaborative robot	Dual LeRobot SO100/101 (BiARM variant)
LeRobot100-RL	`LeRobot100-RL`	RL-optimized lightweight arm (SO100 variant)	LeRobot SO100 Arm, RL
Panda	`Panda`	Single Franka Emika Panda arm	Franka Panda Arm
Panda-RL	`Panda-RL`	RL-optimized Franka Emika Panda arm	Franka Panda Arm, RL
G1-WBC-Joint	`G1-WBC-Joint`	G1 humanoid with whole-body joint control	Unitree G1 WBC (Joint space)
G1-WBC-Pink	`G1-WBC-Pink`	G1 humanoid with WBC variant using "Pink" gait/control	Unitree G1, Pink WBC Variant
G1-Controller-DecoupledWBC	`G1-Controller-DecoupledWBC`	G1 with decoupled whole-body controller for research	Unitree G1, Decoupled WBC
DoublePiper-Abs	`DoublePiper-Abs`	Dual Agilex-Piper mobile manipulators, absolute control	Agilex-Piper Duo Setup (Abs)
DoublePiper-Rel	`DoublePiper-Rel`	Dual Agilex-Piper mobile manipulators, relative control	Agilex-Piper Duo Setup (Rel)
DoublePiper-RL	`DoublePiper-RL`	RL-optimized dual Agilex-Piper manipulators	Agilex-Piper Duo RL
Piper-Abs	`Piper-Abs`	Agilex-Piper mobile manipulator, absolute control	Agilex-Piper Mobile Manipulator (Abs)
Piper-RL	`Piper-RL`	RL-optimized Agilex-Piper manipulator	Agilex-Piper RL Variant
X7s-Abs	`X7s-Abs`	ARX-X7s arm, absolute control	ARX-X7s Arm (Absolute)
X7s-Rel	`X7s-Rel`	ARX-X7s arm, relative control	ARX-X7s Arm (Relative)
X7s-Joint	`X7s-Joint`	ARX-X7s arm, joint position control	ARX-X7s Arm (Joint Space)

Detailed Robot Descriptions

1. Unitree G1 Series

G1-Hand (`--robot=G1-Hand`)

Environment ID: Robocasa-Robot-G1-Hand

Overview:
Full-featured humanoid kitchen robot with dual arm and mobile base, supporting advanced hand and arm control. Designed for human-like kitchen manipulation.

Key Features:

Bimanual manipulation: Differential inverse kinematics for each arm
Handedness: Individual control of dexterous left and right hands
Mobile base: 3-DOF (x, y, yaw) for workspace traversal
High DOF: Flexible for complex and multi-step tasks

Action Space:

base_action: RelativeJointPositionAction      # Mobile base (x, y, yaw, incremental)
  - joint_names: ["base_x_joint", "base_y_joint", "base_yaw_joint"]
  - scale: {"base_x_joint": 0.01, "base_y_joint": 0.01, "base_yaw_joint": 0.02}
  - Control Type: Incremental/local-cartesian

left_arm_action: DifferentialInverseKinematics  # Cartesian delta for left arm
  - target_frame: "tool_left_arm"
  - Control Type: 6-DOF delta pose (pos+ori) in robot base
  - IK: Real-time diff IK, workspace-aware

right_arm_action: DifferentialInverseKinematics # Cartesian delta for right arm
  - target_frame: "tool_right_arm"
  - Control Type: 6-DOF delta pose (pos+ori) in robot base
  - IK: Real-time diff IK, workspace-aware

left_hand_action: HandControl                  # Left hand: multi-finger
  - Control Type: Real hand, fully articulated (multiple fingers)
  - Modes: Grasp, pinch, custom pose possible

right_hand_action: HandControl                 # Right hand: multi-finger
  - Control Type: Real hand, fully articulated (multiple fingers)
  - Modes: Grasp, pinch, custom pose possible

Technical Specifications:

Robot scale: 1.0
Base offset: [0.0, 0.0, 0.8] m
Update/control frequency: 20Hz

G1-Controller (`--robot=G1-Controller`)

Environment ID: Robocasa-Robot-G1-Controller

Overview:
Operator-centric version for teleoperation or joystick-based direct control. Focuses on intuitive base and arm movement, with advanced base-lock feature for precision manipulation.

Key Features:

Direct control: Real-time user or teleoperation mapping; no RL simplification
Base auto-lock: When base joystick held steady, base pose is locked using PID, arms move relative to locked base
PID-based base holding: Ensures stable base while manipulating objects
Flexible hand modes: Swappable hand modes (handle, tracking, etc.)

Action Space:

base_action: RelativeJointPositionAction (with PID lock/unlock)
  - joint_names: ["base_x_joint", "base_y_joint", "base_yaw_joint"]
  - scale: {"base_x_joint": 0.01, "base_y_joint": 0.01, "base_yaw_joint": 0.02}
  - Control Type: Incremental (unlocked) or PID pose-keeping (locked)

left_arm_action: DifferentialInverseKinematics  # or Absolute Pose
  - target_frame: "tool_left_arm"
  - Control Type: 6-DOF delta pose (standard), or absolute pose (when supported)

right_arm_action: DifferentialInverseKinematics # or Absolute Pose
  - target_frame: "tool_right_arm"
  - Control Type: 6-DOF delta pose (standard), or absolute pose (when supported)

left_hand_action: HandControl or Handle Mode

right_hand_action: HandControl or Handle Mode

Special Features:

Auto base lock: Base locks automatically when input is steady, unlocks with movement, using PID to maintain pose
Base-control deadzones: Fine-tuned threshold to prevent drift/jitter
Full teleop mapping: Joystick/teleop control including hand "handle" mode

G1-Loco-Hand (`--robot=G1-Loco-Hand`)

Environment ID: Robocasa-Robot-G1-Loco-Hand

Overview:
Full-body humanoid that enables walking, stepping, and dynamic balancing, suitable for extended workspace and mobility research.

Key Features:

Full locomotion: 12-DOF leg control for walking, squatting, leaning
Multi-modal base: Gait patterns selectable (walk/step, squat modes, etc.)
Dual arms and dexterous hands: Same as G1-Hand
Stable postures: Height, pitch, and roll of base can be adjusted

Action Space:

base_action: LegPositionAction                # Locomotion: 12-DOF (6/leg) + postural targets
  - joint_names: Hip, knee, ankle for both legs
  - Gait/squat config modes

left_arm_action: DifferentialInverseKinematics
  - target_frame: "tool_left_arm"
  - Control Type: 6-DOF delta pose (robust to base movement)

right_arm_action: DifferentialInverseKinematics
  - target_frame: "tool_right_arm"
  - Control Type: 6-DOF delta pose (robust to base movement)

left_hand_action: HandControl (tracking mode; compensation for body motion)
right_hand_action: HandControl (tracking mode; compensation for body motion)

Extra:

Locomotion action includes additional flag/features for "walk", "squat", "balanced stance", etc.

G1-Loco-Controller (`--robot=G1-Loco-Controller`)

Environment ID: Robocasa-Robot-G1-Loco-Controller

Overview:
Teleoperation/joystick variant for full-body humanoid, providing detailed control over walking, posture, arms, and hands. Features a mode switch between locomotion postures (walk, squat, etc.) and finely tuned dead-zones for base orientation control.

Key Features:

Manual gait and base posture control: Selectable via mode/joystick, including height, pitch, roll, yaw
Full-body teleop: Concurrent control of base (including height/orientation), both arms, and hands
Dead-zone logic: Joystick-deadzones prevent jitter in orientation/height (pitch/roll/yaw/height)
Handle/hand modes: Supports handle/teleop and tracking hand control

Action Space:

base_action: LegPositionAction + postural controls
  - [x_local, y_local, yaw_local, height, roll, pitch, yaw] (all normalized, with dead-zone FSM)
  - Modes: standing, walking, squatting, etc.

left_arm_action: DifferentialInverseKinematics or AbsolutePose
  - 6-DOF pose/delta or absolute, as above; locked to base or mobile

right_arm_action: DifferentialInverseKinematics or AbsolutePose
  - 6-DOF pose/delta or absolute, as above; locked to base or mobile

left_hand_action: HandControl (handle or articulated)
right_hand_action: HandControl (handle or articulated)

Special Features:

Base posture dead-zone: Postural axes (roll, pitch, yaw, height) use finite state machine (FSM) to prevent accidental drift
Direct and mode-based base control: Combination of linear (x, y), rotational (yaw), and postural (roll, pitch, height) axes, all user-controllable
Real-time teleop: Designed for direct mapping to gamepad/joystick

G1-RL (`--robot=G1-RL`)

Environment ID: Robocasa-Robot-G1-RL

Overview:
Streamlined single-arm manipulation G1 for efficient reinforcement learning. Prioritizes robot learning efficiency via a minimal action/observation space.

Key Features:

Single right arm only: No left arm or base
Joint-level control: Direct right arm joint angles
Binary right gripper: Fast, simple RL-suitable control
Contact-based feedback: For reward functions and generalization
Minimal complexity: Optimized for accelerated training

Action Space:

right_arm_action: JointPositionAction    # Joint-level for right arm
  - joint_names: Right arm joints (shoulder, elbow, wrist)
  - Control Type: Direct joint angles

right_hand_action: BinaryJointPosition   # Simple open/close gripper
  - joint_names: ["right_hand_*"]
  - 0 = closed, 1 = open

RL Optimizations:

Fast step time, minimal obs
Simple reward computation
Designed for high-throughput RL

G1-FullHand (`--robot=G1-FullHand`)

Environment ID: Robocasa-Robot-G1-FullHand

Overview:
Adds left arm/hand for dual-arm RL from G1-RL base—intended for research in bimanual RL with full finger control.

Key Features:

Both arms controllable: Joint-level policy actions for both arms (all joints)
Both grippers/hands: Full or binary hand action per arm

Action Space:

left_arm_action: JointPositionAction
  - joint_names: All left arm and hand joints
right_arm_action: JointPositionAction
  - joint_names: All right arm joints
left_hand_action: HandControl or BinaryJointPosition (per env)
right_hand_action: HandControl or BinaryJointPosition (per env)

G1-Controller-DecoupledWBC (`--robot=G1-Controller-DecoupledWBC`)

Environment ID: Robocasa-Robot-G1-Controller-DecoupledWBC

Overview:
This configuration augments the Unitree G1 robot with a Decoupled Whole-Body Controller (WBC), enabling advanced coordinated control over both the base and manipulators for complex whole-body tasks.

Key Features:

Decoupled Whole-Body Control: Independent high-level control of base and both arms for complex motion or manipulation tasks.
Full 7-DOF arms: Both arms are actuated, with flexible combinations of position and velocity control.
Customizable hand control: Supports multiple hand/gripper action schemes.
Extended base control: Includes not just base movement but also torso orientation and height, suitable for dynamic and versatile scenarios.
Specialized dead-zone logic: For fine-grained manipulation of base and orientation via gamepad inputs or high-level policy control.

Action Space:

base_action: DecoupledWBCBaseAction
  - Controls: [lin_x_local, lin_y_local, ang_z_local, base_height_cmd, torso_roll, torso_pitch, torso_yaw]
  - Note: Specialized logic enables stand/squat, turning, and orientation control based on input mode.

left_arm_action: DecoupledWBCKinematicArmAction or similar
  - Joint-level or Cartesian control over all left arm joints.

right_arm_action: DecoupledWBCKinematicArmAction or similar
  - Joint-level or Cartesian control over all right arm joints.

left_hand_action: Gripper/HandControl (selectable modes, e.g., handle, tracking)
right_hand_action: Gripper/HandControl (selectable modes, e.g., handle, tracking)

RL and Teleop Notes:

Base locking and PID: The base controller includes auto-locking PID loops for stability when no user or policy input is given, unlocking when input magnitude rises.
Dead zone/activation logic: The base and torso orientation axes include dead-zone logic for safe, intuitive teleop, reducing accidental drift.
Multi-modal control: Switch easily between teleoperation and RL control, including coordinated whole-body actions.

Typical Use Cases:

Whole-body mobile manipulation
Teleop with advanced base and torso motions
RL tasks requiring fine control and stability across locomotion and manipulation

2. PandaOmron Mobile Manipulator Series

PandaOmron-Rel (`--robot=PandaOmron-Rel`)

Environment ID: Robocasa-Robot-PandaOmron-Rel

Overview: A mobile manipulation platform combining the precision of Franka Panda arm with Omron mobile base mobility. Uses relative position control for smooth, incremental movements.

Key Features:

7-DOF manipulation: Franka Panda arm with high precision
4-DOF mobile base: Forward/side/yaw/height adjustment
Relative control: Incremental position commands
High precision: Sub-millimeter accuracy for manipulation

Action Space:

arm_action: DifferentialInverseKinematics  # Arm Cartesian control
  - target_frame: "panda_hand" (end-effector)
  - Control Type: 6-DOF Cartesian space incremental control
  - IK solver: Real-time differential IK with collision avoidance
  - Workspace: 855mm reach radius with full orientation control
  - Precision: Sub-millimeter positioning accuracy
  - Joint limits: Full 7-DOF joint range with safety limits

gripper_action: BinaryJointPosition       # Gripper open/close
  - joint_names: ["panda_finger_joint.*"]
  - Control Type: Binary parallel gripper control
  - Open command: 0.04m finger separation
  - Close command: 0.0m (contact-based closing)
  - Force limit: 200N maximum grip force
  - Speed: 0.2m/s maximum finger velocity

base_action: RelativeJointPosition        # Mobile base control
  - mobilebase_forward: Forward/backward incremental movement
    - Scale: Small increments for smooth motion
    - Range: Continuous movement within workspace bounds
  - mobilebase_side: Left/right lateral movement
    - Scale: Small increments for precise positioning
    - Range: Full kitchen workspace coverage
  - mobilebase_yaw: Rotational movement
    - Scale: Small angular increments
    - Range: 360° rotation capability
  - mobilebase_torso_height: Vertical adjustment
    - Scale: Small height increments
    - Range: Variable working height for different tasks

PandaOmron-Abs (`--robot=PandaOmron-Abs`)

Environment ID: Robocasa-Robot-PandaOmron-Abs

Overview: Same hardware as PandaOmron-Rel but uses absolute position control for precise, goal-directed movements. Better suited for tasks requiring exact positioning.

Key Features:

Absolute positioning: Direct coordinate targeting
Same hardware: Identical to Rel version
Precise control: Better for waypoint navigation
Goal-oriented: Suitable for planned motion sequences

Action Space:

arm_action: DifferentialInverseKinematics  # Arm Cartesian control
  - Control Type: Absolute 6-DOF Cartesian targeting
  - Target specification: Direct world coordinates
  - Motion planning: Trajectory optimization to target
  - Precision: Direct goal reaching with path planning

gripper_action: BinaryJointPosition       # Same as Rel version
base_action: RelativeJointPosition        # Base control
  - Control Type: Absolute position targeting for base
  - Goal specification: Direct world coordinates for base pose
  - Motion planning: Optimal path planning to target location

Use Cases:

Precise pick-and-place operations
Waypoint-based navigation
Calibrated manipulation tasks
Multi-step planned sequences

3. DoublePanda Dual-Arm Series

DoublePanda-Rel (`--robot=DoublePanda-Rel`)

Environment ID: Robocasa-Robot-DoublePanda-Rel

Overview: Sophisticated dual-arm manipulation system using two Franka Panda robots in a fixed configuration. Enables complex bimanual manipulation tasks with relative position control.

Key Features:

Dual 7-DOF arms: Two synchronized Franka Panda arms
Independent control: Each arm controlled separately
Coordinated manipulation: Can work together on single objects
Symmetric design: Mirrored arm configuration
Contact feedback: Sensors on both arms

Action Space:

left_arm_action: DifferentialInverseKinematics   # Left arm control
  - target_frame: "ee_tcp_L" (left end-effector)
  - Control Type: 6-DOF Cartesian incremental control
  - joint_names: ["panda_L_joint[1-7]"]
  - Workspace: Full 855mm reach with orientation control
  - Coordination: Independent motion with collision avoidance
  - Precision: Sub-millimeter positioning for precise tasks

right_arm_action: DifferentialInverseKinematics  # Right arm control
  - target_frame: "ee_tcp_R" (right end-effector)
  - Control Type: 6-DOF Cartesian incremental control
  - joint_names: ["panda_R_joint[1-7]"]
  - Workspace: Full 855mm reach with orientation control
  - Coordination: Independent motion with collision avoidance
  - Precision: Sub-millimeter positioning for precise tasks

left_gripper_action: BinaryJointPosition        # Left gripper
  - joint_names: ["panda_L_finger.*"]
  - Control Type: Binary parallel gripper
  - Open command: 0.04m finger separation
  - Close command: 0.0m (contact-based)
  - Force control: 200N maximum grip force
  - Coordination: Can coordinate with right gripper for bimanual grasping

right_gripper_action: BinaryJointPosition       # Right gripper
  - joint_names: ["panda_R_finger.*"]
  - Control Type: Binary parallel gripper
  - Open command: 0.04m finger separation
  - Close command: 0.0m (contact-based)
  - Force control: 200N maximum grip force
  - Coordination: Can coordinate with left gripper for bimanual grasping

Arm Configuration:

Left arm offset: [0.389, 0.2, 0.4658] meters
Right arm offset: [0.389, -0.2, 0.4658] meters
Joint naming: panda_L_joint[1-7] and panda_R_joint[1-7]

DoublePanda-Abs (`--robot=DoublePanda-Abs`)

Environment ID: Robocasa-Robot-DoublePanda-Abs

Overview: Same dual-arm hardware as DoublePanda-Rel but with absolute position control for precise, coordinated bimanual manipulation.

Key Features:

Absolute coordination: Precise relative positioning between arms
Synchronized motion: Coordinated dual-arm trajectories
High precision: Sub-millimeter coordination accuracy
Fixed base: Stable platform for precise manipulation

Action Space:

left_arm_action: DifferentialInverseKinematics   # Left arm control
  - Control Type: Absolute 6-DOF Cartesian targeting
  - Target specification: Direct world coordinates
  - Motion planning: Coordinated trajectory with right arm
  - Collision avoidance: Inter-arm collision prevention
  - Synchronization: Temporal coordination with right arm motion

right_arm_action: DifferentialInverseKinematics  # Right arm control
  - Control Type: Absolute 6-DOF Cartesian targeting
  - Target specification: Direct world coordinates
  - Motion planning: Coordinated trajectory with left arm
  - Collision avoidance: Inter-arm collision prevention
  - Synchronization: Temporal coordination with left arm motion

left_gripper_action: BinaryJointPosition        # Left gripper
  - Coordination: Synchronized grasping with right gripper
  - Force distribution: Coordinated force control for bimanual tasks

right_gripper_action: BinaryJointPosition       # Right gripper
  - Coordination: Synchronized grasping with left gripper
  - Force distribution: Coordinated force control for bimanual tasks

Use Cases:

Bimanual assembly tasks
Large object manipulation
Coordinated pick-and-place
Complex kitchen operations (e.g., opening containers while pouring)

4. Agilex Piper Series

Piper-Abs / DoublePiper-Abs (`--robot=Piper-Abs`, `--robot=DoublePiper-Abs`)

Environment ID:

Robocasa-Robot-Piper-Abs
Robocasa-Robot-DoublePiper-Abs

Overview:
The Agilex Piper is a mobile manipulator platform featuring a 6-DOF industrial manipulator mounted on an AGV base. The Abs variants use absolute pose/joint control for both the manipulator and gripper, providing precise, repeatable motions ideal for industrial-style tasks.

Key Features:

Mobile base: Differential drive or holonomic options for workspace mobility.
6-DOF arm: Supports high-precision positioning and kinematic operations.
Eye-in-hand camera support for real-time perception.
Absolute joint control: Direct mapping of observations/commands to arm joints.
Pinocchio/analytical IK support for pose-to-joint conversion.

Action Space:

# For Piper-Abs (single-arm)
arm_action: JointPositionAction             # Absolute joint positions for manipulator arm
  - joint_names: ["joint1", "joint2", "joint3", "joint5", "joint6"]
  - scale: 1.0
  - Control Type: Joint target positions (typically 5-DOF subset)

gripper_action: BinaryJointPosition         # Open/close gripper
  - joint_names: ["finger_joint.*"]
  - open_command_expr: {"finger_joint_left": 0.035, "finger_joint_right": -0.035}
  - close_command_expr: {"finger_joint.*": 0.0}

For DoublePiper-Abs, actions are duplicated for left and right arms/grippers.

Technical Specifications:

Robot scale: 1.0 (configurable)
End-effector offset: configurable in kinematic chain
Frequency: 20Hz control loop
Pinocchio IK: Supported for absolute pose control inputs

Use Cases:

Precise pick-and-place or manipulation
Workspace coverage with mobile base
Teleoperation, imitation, or policy execution with absolute pose

Piper-RL / DoublePiper-RL (`--robot=Piper-RL`, `--robot=DoublePiper-RL`)

Environment ID:

Robocasa-Robot-Piper-RL
Robocasa-Robot-DoublePiper-RL

Overview:
RL-optimized variants streamlining the action/observation space for efficient reinforcement learning, using joint-space control and simplified gripper logic.

Key Features:

Joint position control: RL agent outputs normalized or absolute arm joint positions.
Binary gripper action: RL agent can open/close gripper with a single command dimension.
Fast & robust simulation: Minimalist sensing, simplified base physics.

Action Space:

# For Piper-RL (single-arm)
arm_action: JointPositionAction
  - joint_names: ["joint1", "joint2", "joint3", "joint5", "joint6"]
gripper_action: BinaryJointPosition
  - joint_names: ["finger_joint.*"]

For DoublePiper-RL, actions are duplicated for both arms.

Technical Specifications:

5-DOF arm control (joint1, joint2, joint3, joint5, joint6)
Frequency: 20Hz
Recommended for high-throughput RL experiments

Use Cases:

RL policy learning on mobile manipulators
Multi-arm RL research and transfer learning

5. ARX X7s Series

X7s-Abs / X7s-Rel / X7s-Joint (`--robot=X7s-Abs`, `--robot=X7s-Rel`, `--robot=X7s-Joint`)

Environment ID:

Robocasa-Robot-X7s-Abs
Robocasa-Robot-X7s-Rel
Robocasa-Robot-X7s-Joint

Overview:
The ARX X7s series includes a compact, 7-DOF dual-arm robot platform with numerous control modes for tabletop and bimanual kitchen manipulation. X7s variants emphasize high modularity: absolute (cartesian), relative (incremental), and direct joint-space arm control are all supported.

Key Features:

Bimanual 7-DOF arms: Each arm is independently actuated with full 7-joint kinematic chains.
Flexible action spaces: Choose absolute (cartesian pose), relative (delta pose), or direct joint angles (joint position).
End-effector cameras: Supports multiple wrist/hand/shoulder camera viewpoints for vision-based tasks.
Gripper options: Independent binary grippers for each arm.
Optional Pinocchio IK: Fast analytical inverse kinematics for absolute pose tracking.

Action Space:

X7s-Abs (absolute pose control, Pinocchio IK)

left_arm_action: JointPositionAction
  - joint_names: [
      "left_shoulder_y", "left_shoulder_x", "left_shoulder_z",
      "left_elbow_y", "left_elbow_x",
      "left_wrist_y", "left_wrist_z"
    ]
  - Control Type: low-level joint command from Pinocchio IK solution (input: desired left hand pose)
right_arm_action: JointPositionAction
  - joint_names: [
      "right_shoulder_y", "right_shoulder_x", "right_shoulder_z",
      "right_elbow_y", "right_elbow_x",
      "right_wrist_y", "right_wrist_z"
    ]
  - Control Type: low-level joint command from Pinocchio IK solution (input: desired right hand pose)
left_gripper_action: BinaryJointPosition
  - joint_names: ["left_gripper1", "left_gripper2"]
right_gripper_action: BinaryJointPosition
  - joint_names: ["right_gripper1", "right_gripper2"]
base_action: RelativeJointPositionAction  # Optional: for holonomic base, if equipped
body_action: RelativeJointPositionAction  # Optional: for z/y prismatic body

X7s-Rel (relative pose/delta control, differential IK)

left_arm_action: DifferentialInverseKinematics
  - joint_names: ["left_shoulder.*", "left_wrist.*", "left_elbow.*"]
  - target_frame: "left_hand_link"
  - Control Type: delta pose (6-DOF) in base_link frame
right_arm_action: DifferentialInverseKinematics
  - joint_names: ["right_shoulder.*", "right_wrist.*", "right_elbow.*"]
  - target_frame: "right_hand_link"
  - Control Type: delta pose (6-DOF) in base_link frame
...
# Gripper and base as above

X7s-Joint (direct joint-space control)

left_arm_action: JointPositionAction
  - joint_names: [
      "left_shoulder_y", "left_shoulder_x", "left_shoulder_z",
      "left_elbow_y", "left_elbow_x",
      "left_wrist_y", "left_wrist_z"
    ]
right_arm_action: JointPositionAction
  - joint_names: [
      "right_shoulder_y", "right_shoulder_x", "right_shoulder_z",
      "right_elbow_y", "right_elbow_x",
      "right_wrist_y", "right_wrist_z"
    ]
# Gripper and base as above

Cameras (available views for all variants):

Left hand camera: {ENV_REGEX_NS}/Robot/left_hand_link/left_hand_camera
Right hand camera: {ENV_REGEX_NS}/Robot/right_hand_link/right_hand_camera
Eye-in-hand & first person: configurable
Shoulder cameras: {ENV_REGEX_NS}/Robot/link1/[left/right]_shoulder_camera

Technical Specifications:

Robot scale: 1.0 (configurable)
7-DOF per arm, rigid base or optional holonomic base
End-effector offset: configurable per use case (default ~0.17m palm offset)
Control frequency: 20Hz (configurable)
Supported with fast Pinocchio IK (for absolute pose control)

Use Cases:

Bimanual kitchen manipulation & coordinated reaching
Tabletop fine motor skills testing
RL or traditional control with full and incremental control modes
Visual manipulation experiments with integrated end-effector cameras

6. Franka Series

Franka (`--robot=Panda`)

Environment ID: Robocasa-Robot-Panda

Overview:
A single Franka Emika Panda arm with high-precision 7-DOF manipulation and parallel jaw gripper, ideal for tabletop and single-arm manipulation tasks.

Key Features:

7-DOF arm: Full joint range, high-precision, fast response
Versatile gripper: Parallel finger control
Flexible control: Supports both differential IK and direct joint commands
Small scale: Suited for diverse lab or simulated environments

Action Space:

arm_action: DifferentialInverseKinematics   # 6-DOF relative (delta pose) or absolute pose control
  - target_frame: "panda_hand" (TCP)
  - Control Type: 
      - Relative mode (default): delta pose in robot base
      - Absolute mode (for Abs variant): full 6-DOF target pose
  - IK solver: Damped least squares (DLS)
  - joint_names: ["panda_joint.*"]

gripper_action: BinaryJointPosition
  - joint_names: ["panda_finger.*"]
  - open_command_expr: {"panda_finger_.*": 0.04}
  - close_command_expr: {"panda_finger_.*": 0.0}

Franka-Abs (`--robot=Panda-Abs`)

Environment ID: Robocasa-Robot-Panda-Abs

Overview:
Absolute pose (Cartesian) control over the same Franka Panda arm; ideal for precise “move-to-pose” and goal-reaching tasks where exact coordinate targeting is required.

Key Features:

Absolute 6-DOF targeting: Directly specify tool pose in world or base frame
Trajectory optimization: Path planning to target pose
High precision: Suited for pick/place, assembly, or calibrated actions

Action Space:

arm_action: DifferentialInverseKinematics
  - Control Type: Absolute 6-DOF Cartesian pose
  - joint_names: ["panda_joint.*"]
  - body_name: "panda_hand"
  - controller: DifferentialIKControllerCfg(command_type="pose", use_relative_mode=False, ik_method="dls")
  - body_offset: (0.0, 0.0, 0.107)

gripper_action: BinaryJointPosition
  - As above

Franka-RL (`--robot=Panda-RL`)

Environment ID: Robocasa-Robot-Panda-RL

Overview:
RL-optimized control variant of the Franka Panda, designed for reinforcement learning with direct joint position control and streamlined observation spaces for policy training.

Key Features:

Direct joint command: Simple, fast policy learning
Full 7-DOF joint control: All Panda arm joints exposed to policy
Binary gripper: Open/close action for rapid reward feedback
Compatible with: Gym / IsaacLab RL pipelines

Action Space:

arm_action: JointPositionAction
  - joint_names: ["panda_joint.*"]
  - scale: 1.0
  - use_default_offset: True

gripper_action: BinaryJointPosition
  - As above

Use Cases:

Tabletop pick-and-place
RL skill learning and benchmarking
Precise single-arm manipulation and calibration
Vision-based manipulation (with optional cameras)

7. Arena G1 Series

G1-Arena-Joint / G1-Arena-Pink (`--robot=G1-Arena-Joint`, `--robot=G1-Arena-Pink`)

Environment ID:

Robocasa-Robot-G1-Arena-Joint
Robocasa-Robot-G1-Arena-Pink

Overview:
The Arena G1 Series leverages the IsaacLab-Arena built-in G1 humanoid robot platform, featuring a whole-body controller (WBC) and joint-level control for advanced research and benchmarking. Ideal for full-body humanoid manipulation, locomotion, and kitchen tasks with modular action spaces.

Key Features:

Full-body WBC: Decoupled whole-body controller for coordinating base, torso, and dual arms.
Flexible action composition: Joint-level and WBC-pink variant support, matching real robot structure.
Camera-ready: Supports camera-enabled observations for vision-based policy learning.
Synchronized dual arms & hands: All arm and hand joints are controllable, supporting complex bimanual tasks.
Precise mobile base control: Local x/y/yaw, base height, and torso orientation available.
Ready for RL and teleop: Standardized observation and action formats.

Action Space:

base_action: DecoupledWBCBaseAction
  - Controls: [lin_x_local, lin_y_local, ang_z_local, base_height_cmd, torso_roll, torso_pitch, torso_yaw]
  - Range: Each value in [-1, 1]; physical scaling handled internally.

left_arm_action: JointPositionAction
  - joint_names: All left arm joints (see G1 documentation)

right_arm_action: JointPositionAction
  - joint_names: All right arm joints (see G1 documentation)

left_hand_action: HandControl or BinaryJointPosition
right_hand_action: HandControl or BinaryJointPosition

For G1-Arena-Pink, the action/observation spaces are adapted for the "Pink" gait/control variant for research.

Technical Details:

Initial robot base height: 0.75 m (configurable)
Flexible joint/torso orientation and dual-arm control
Compatible with IsaacLab-Arena and LW-BenchHub multitask scenarios
All actions can be provided as range [-1, 1], internally mapped to physical robot

Use Cases:

Full-body kitchen manipulation and navigation
Complex teleoperation with decoupled base/arm/torso
RL research using modular humanoid control
Dual-arm coordinated grasping, cooking, or assistive tasks
Vision-based skill learning using built-in cameras

8. LeRobot Series

LeRobot-RL / LeRobot100-RL / LeRobot-AbsJointGripper-RL / LeRobot-BiARM-RL

Environment IDs:

Robocasa-Robot-LeRobot-RL
Robocasa-Robot-LeRobot100-RL
Robocasa-Robot-LeRobot-AbsJointGripper-RL
Robocasa-Robot-LeRobot-BiARM-RL

Overview:
The LeRobot series provides affordable, modular robot arms for RL and imitation learning research, available in single-arm and dual-arm (BiARM) variants. These robots offer flexible, simplified control for effective table-top and workspace manipulation. The RL variants emphasize focused action/observation spaces and rapid simulation.

Key Features:

Single- or dual-arm options: Choose from LeRobot-RL (SO101/SO100 arm), or BiARM RL with synchronized bimanual workspace.
Gripper support: Both continuous and absolute joint gripper control available; AbsJointGripper variant for task-specific gripper policies.
End-effector cameras: Integrated hand and global cameras for learning from observation.
Affordable/accessible hardware: Designed for reproducible research and rapid prototyping.
Contact sensors: Detect gripper/table and gripper/object contact.

Action Space:

LeRobot-RL (SO101, single arm, relative joint increments)

arm_action: RelJointPositionAction
  - joint_names: ["shoulder.*", "elbow_flex", "wrist.*", "gripper"]
  - scale: { "shoulder.*": 0.05, "elbow_flex": 0.05, "wrist.*": 0.05, "gripper": 0.2 }
  - use_zero_offset: True
  - clip: { all: [-1, 1] }

LeRobot-AbsJointGripper-RL (absolute joint command + limit for gripper)

arm_action: JointPositionAction
  - joint_names: ["shoulder.*", "elbow_flex", "wrist.*"]
  - scale: 1
  - use_default_offset: False

gripper_action: JointPositionLimitAction
  - joint_names: ["gripper"]
  - scale: 1
  - use_default_offset: False

LeRobot100-RL (SO100 shorter arm, similar to above)

arm_action: RelJointPositionAction
  - joint_names: ["shoulder.*", "elbow_flex", "wrist.*", "gripper"]

LeRobot-BiARM-RL (dual-arm, absolute joints)

left_arm_action: JointPositionAction
  - joint_names: ["left_shoulder.*", "left_elbow_flex", "left_wrist.*"]
left_gripper_action: JointPositionAction
  - joint_names: ["left_gripper"]

right_arm_action: JointPositionAction
  - joint_names: ["right_shoulder.*", "right_elbow_flex", "right_wrist.*"]
right_gripper_action: JointPositionAction
  - joint_names: ["right_gripper"]

Technical Specifications:

6-DOF (SO101) or 5-DOF (SO100) per arm
Supports table-top and mobile experimental setups
20Hz control loop
Reward/focus on gripper fingertips for manipulation policies

Use Cases:

Efficient RL/IL on accessible hardware
Grasping, pick-and-place, basic tool use
Imitation learning, contact-rich skills with touch/camera sensing
Bimanual skill learning with the BiARM variant

Robot Selection Guidelines

Task Type	Recommended Robot	Reason/Justification
Complex full-body tasks	G1-Hand	Humanoid with mobility and dexterous hands for versatile tasks
Mobile manipulation	PandaOmron-Rel / PandaOmron-Abs	Precision arm on a mobile base for workspace flexibility
Dual-arm (bimanual) manipulation	DoublePanda-Rel / DoublePanda-Abs / X7s-Abs/Rel	Synchronized dual-arm control (DoublePanda: fixed; X7s: mobile base)
Mobile dual-arm platform	X7s-Abs / X7s-Rel / X7s-Joint	Dual 7-DOF arms + optional holonomic base for tabletop or mobile use
RL-focused fixed-arm research	Piper-RL / LeRobot-RL / LeRobot100-RL	Lightweight, fixed-base manipulators well-suited for RL benchmarking
Fixed-base mobile manipulator	Piper-Abs / Piper-RL	Agilex-Piper: compact arm for fixed workstation or mobile tasks
Compact & modular arms	LeRobot-RL / LeRobot-BiARM-RL	Modular design for scaling single to dual-arm experiments
Reinforcement learning	G1-RL / Piper-RL / LeRobot-RL	Simplified control and fast simulation for RL policy development
Locomotion & mobility	G1-Loco-Hand	Advanced base mobility with arm/leg coordination
High-precision tasks	PandaOmron-Abs	High-precision Cartesian and joint control, sub-millimeter accuracy

This comprehensive LW-BenchHub robot suite enables researchers to select the most suitable platform for diverse manipulation research needs—spanning teleoperation, RL, precision tasks, bimanual and mobile manipulation, on both fixed and mobile base setups.

Robot Configuration Summary Table​

Detailed Robot Descriptions​

1. Unitree G1 Series​

G1-Hand (--robot=G1-Hand)​

G1-Controller (--robot=G1-Controller)​

G1-Loco-Hand (--robot=G1-Loco-Hand)​

G1-Loco-Controller (--robot=G1-Loco-Controller)​

G1-RL (--robot=G1-RL)​

G1-FullHand (--robot=G1-FullHand)​

G1-Controller-DecoupledWBC (--robot=G1-Controller-DecoupledWBC)​

2. PandaOmron Mobile Manipulator Series​

PandaOmron-Rel (--robot=PandaOmron-Rel)​

PandaOmron-Abs (--robot=PandaOmron-Abs)​

3. DoublePanda Dual-Arm Series​

DoublePanda-Rel (--robot=DoublePanda-Rel)​

DoublePanda-Abs (--robot=DoublePanda-Abs)​

4. Agilex Piper Series​

Piper-Abs / DoublePiper-Abs (--robot=Piper-Abs, --robot=DoublePiper-Abs)​

Piper-RL / DoublePiper-RL (--robot=Piper-RL, --robot=DoublePiper-RL)​

5. ARX X7s Series​

X7s-Abs / X7s-Rel / X7s-Joint (--robot=X7s-Abs, --robot=X7s-Rel, --robot=X7s-Joint)​

X7s-Abs (absolute pose control, Pinocchio IK)​

X7s-Rel (relative pose/delta control, differential IK)​

X7s-Joint (direct joint-space control)​

6. Franka Series​

Franka (--robot=Panda)​

Franka-Abs (--robot=Panda-Abs)​

Franka-RL (--robot=Panda-RL)​

7. Arena G1 Series​

G1-Arena-Joint / G1-Arena-Pink (--robot=G1-Arena-Joint, --robot=G1-Arena-Pink)​

8. LeRobot Series​

LeRobot-RL / LeRobot100-RL / LeRobot-AbsJointGripper-RL / LeRobot-BiARM-RL​

LeRobot-RL (SO101, single arm, relative joint increments)​

LeRobot-AbsJointGripper-RL (absolute joint command + limit for gripper)​

LeRobot100-RL (SO100 shorter arm, similar to above)​

LeRobot-BiARM-RL (dual-arm, absolute joints)​

Robot Selection Guidelines​

Robot Configuration Summary Table

Detailed Robot Descriptions

1. Unitree G1 Series

G1-Hand (`--robot=G1-Hand`)

G1-Controller (`--robot=G1-Controller`)

G1-Loco-Hand (`--robot=G1-Loco-Hand`)

G1-Loco-Controller (`--robot=G1-Loco-Controller`)

G1-RL (`--robot=G1-RL`)

G1-FullHand (`--robot=G1-FullHand`)

G1-Controller-DecoupledWBC (`--robot=G1-Controller-DecoupledWBC`)

2. PandaOmron Mobile Manipulator Series

PandaOmron-Rel (`--robot=PandaOmron-Rel`)

PandaOmron-Abs (`--robot=PandaOmron-Abs`)

3. DoublePanda Dual-Arm Series

DoublePanda-Rel (`--robot=DoublePanda-Rel`)

DoublePanda-Abs (`--robot=DoublePanda-Abs`)

4. Agilex Piper Series

Piper-Abs / DoublePiper-Abs (`--robot=Piper-Abs`, `--robot=DoublePiper-Abs`)

Piper-RL / DoublePiper-RL (`--robot=Piper-RL`, `--robot=DoublePiper-RL`)

5. ARX X7s Series

X7s-Abs / X7s-Rel / X7s-Joint (`--robot=X7s-Abs`, `--robot=X7s-Rel`, `--robot=X7s-Joint`)

X7s-Abs (absolute pose control, Pinocchio IK)

X7s-Rel (relative pose/delta control, differential IK)

X7s-Joint (direct joint-space control)

6. Franka Series

Franka (`--robot=Panda`)

Franka-Abs (`--robot=Panda-Abs`)

Franka-RL (`--robot=Panda-RL`)

7. Arena G1 Series

G1-Arena-Joint / G1-Arena-Pink (`--robot=G1-Arena-Joint`, `--robot=G1-Arena-Pink`)

8. LeRobot Series

LeRobot-RL / LeRobot100-RL / LeRobot-AbsJointGripper-RL / LeRobot-BiARM-RL

LeRobot-RL (SO101, single arm, relative joint increments)

LeRobot-AbsJointGripper-RL (absolute joint command + limit for gripper)

LeRobot100-RL (SO100 shorter arm, similar to above)

LeRobot-BiARM-RL (dual-arm, absolute joints)

Robot Selection Guidelines