Skip to content

README.md

Latest commit

 

History

History
106 lines (79 loc) · 5.45 KB

README.md

File metadata and controls

106 lines (79 loc) · 5.45 KB
Raw

Built-in Robots

This is the canonical list of robot embodiments that ship with RoboLab. For how to use a robot (registration wiring, defining your own robot, contact grippers, wrist cameras), see docs/robots.md.

Robot Embodiment Action spaces Cameras
DROID
(Franka + Robotiq 2F-85)
droid.py		 single-arm fixed-base parallel-jaw joint position, absolute EE IK, relative EE IK wrist
Franka Panda
franka.py, franka_high_pd.py		 single-arm fixed-base parallel-jaw joint position, absolute EE IK, relative EE IK —
Kinova Gen3
(Gen3 7-DoF + Robotiq 2F-85)
kinova_gen3.py		 single-arm fixed-base parallel-jaw joint position wrist

Gripper convention for all binary gripper actions: a scalar per gripper, > 0.5 closes, ≤ 0.5 opens. Quaternions are (w, x, y, z); absolute IK targets are expressed in the robot root frame, translations in meters.

DROID (Franka + Robotiq 2F-85)

tags: single-arm · fixed-base · parallel-jaw · wrist-cam · gravity-disabled · high-PD · benchmark-default

The default benchmark embodiment: a Franka Panda arm with a Robotiq 2F-85 gripper, matching the DROID platform. High PD gains (400/80) with gravity disabled on the arm, plus a 720p wrist camera whose intrinsics are calibrated to match pi05 / DreamZero training data.

Action config Layout Dim
DroidJointPositionActionCfg 7 arm joint targets + binary gripper 8
DroidIKActionCfg absolute EE pose (x, y, z, qw, qx, qy, qz) + binary gripper 8
DroidRelIKActionCfg relative EE pose (dx, dy, dz, droll, dpitch, dyaw) + binary gripper 7
  • Config classes: DroidCfg (robot + wrist camera + EE frame transformers)
  • Proprioception: ProprioceptionObservationCfg — arm joint positions, gripper open fraction, EE pose (both the gripper mount flange ee_* and the rotated control frame eef_*)
  • Contact gripper: {"gripper": ...left_inner_finger}
  • Registrations: robolab/registrations/droid/ (jointpos, abs-IK, rel-IK, lighting/background variations)
from robolab.robots.droid import DroidCfg, DroidJointPositionActionCfg, contact_gripper

Franka Panda

tags: single-arm · fixed-base · parallel-jaw

A stock Franka Panda with its factory finger gripper. Two articulation variants share the same action configs: franka.py with standard PD gains (80/4), and franka_high_pd.py with high gains (400/80) and gravity disabled (better target tracking for policy control).

Action config Layout Dim
FrankaJointPositionActionCfg 7 arm joint targets + binary gripper 8
FrankaIKActionCfg absolute EE pose (x, y, z, qw, qx, qy, qz) + binary gripper 8
FrankaRelIKActionCfg relative EE pose (dx, dy, dz, droll, dpitch, dyaw) + binary gripper 7
  • Config classes: FrankaCfg (one per variant file; action configs in franka_definitions.py)
  • Proprioception: EE frame pose and finger joint positions (franka_definitions.py)
  • Contact gripper: {"gripper": ...panda_leftfinger}
from robolab.robots.franka import FrankaCfg                      # standard PD
from robolab.robots.franka_high_pd import FrankaCfg              # high PD, gravity disabled
from robolab.robots.franka_definitions import FrankaJointPositionActionCfg, contact_gripper

Kinova Gen3 (Gen3 7-DoF + Robotiq 2F-85)

tags: single-arm · fixed-base · parallel-jaw · wrist-cam · gravity-disabled

A Kinova Gen3 7-DoF arm with a Robotiq 2F-85 gripper, welded to the world at the base (fix_root_link=True, gravity disabled on the arm). The USD is vendored under assets/robots/kinova_gen3_robotiq_2f85/ and derived from Kinova ros2_kortex and PickNik ros2_robotiq_gripper at pinned revisions — see the README and license files in that folder.

The gripper's six joints are driven together from one binary action rather than through PhysX mimic constraints; the signed per-joint targets live in GRIPPER_JOINT_COMMANDS.

Action config Layout Dim
KinovaJointPositionActionCfg 7 arm joint targets + binary gripper 8
  • Config classes: KinovaGen3Cfg (robot + wrist camera + EE frame transformer), KinovaWristCameraCfg (exposes the wrist camera to image observations)
  • Proprioception: KinovaProprioceptionObservationCfg — arm joint positions, gripper open fraction, EE pose at robotiq_85_base_link
  • Contact gripper: {"gripper": ...robotiq_85_.*_finger_tip_link}
  • Registrations: robolab/registrations/kinova/ (jointpos)
from robolab.robots.kinova_gen3 import (
    KinovaGen3Cfg,
    KinovaJointPositionActionCfg,
    KinovaProprioceptionObservationCfg,
    contact_gripper,
)

Actuator gains are simulation defaults, not measured from hardware — this is a functional simulation model rather than a calibrated digital twin.

Also in this folder: delta_actions.py, a helper that converts a target EE pose into a relative (delta) pose action — used by trajectory replay, not an action space itself.

The robot stills above are rendered in an empty scene at each robot's reset posture.