Implementation:Google deepmind Dm control Jaco Arm Hand

Metadata
Knowledge Sources	dm_control
Domains	Reinforcement Learning, Robotics Simulation, Robot Modelling
Last Updated	2026-02-15 00:00 GMT

Overview

Concrete tool for configuring the Kinova Jaco 6-DOF arm and 3-finger hand as composer entities with velocity actuators, torque sensors, and manipulation-specific observation defaults.

Description

The dm_control manipulation suite uses the Kinova Jaco robot, modelled as two composer entities:

JacoArm (jaco_arm.py):

• Loads the jaco_arm.xml MJCF model.
• Contains 6 revolute joints split into 3 large joints (joint_1, joint_2, joint_3) and 3 small joints (joint_4, joint_5, joint_6).
• Adds a velocity actuator to each joint with torque and velocity limits from the Kinova datasheet: large joints get a peak torque of 30.5 Nm and max velocity of 36 deg/s; small joints get 6.8 Nm and 48 deg/s.
• Adds a 3-axis torque sensor per joint.
• The JacoArmObservables class returns joint positions as sin/cos pairs (bounded) and projects torque sensor readings onto the joint axis.

JacoHand (jaco_hand.py):

• Loads the jaco_hand.xml MJCF model.
• Contains 3 finger joints, each with a velocity actuator (gain=10, force range +/-1 N, ctrl range +/-5).
• Exposes a tool_center_point property (either the grip site or the pinch site, configurable via use_pinch_site_as_tcp).
• Provides a set_grasp(physics, close_factors) method that positions fingers to a desired open/close fraction (0 = fully open, 1 = fully closed).
• The JacoHandObservables class provides pinch-site position and rotation matrix observables.

Factory functions (robots.py):

• make_arm(obs_settings) -- constructs a JacoArm with the observation options derived from JACO_ARM_OBSERVABLES (proprio: joints_pos, joints_vel; ftt: joints_torque).
• make_hand(obs_settings) -- constructs a JacoHand with use_pinch_site_as_tcp=True and observation options derived from JACO_HAND_OBSERVABLES (proprio: joints_pos, joints_vel, pinch_site_pos, pinch_site_rmat).

The default arm base offset in the arena is (0.0, 0.4, 0.0).

Usage

Task modules call robots.make_arm() and robots.make_hand(), then attach the hand to the arm and the arm to the arena. This is typically done inside the task's factory function (e.g. _reach(), _lift()).

Code Reference

Attribute	Value
Source Locations	dm_control/entities/manipulators/kinova/jaco_arm.py, lines 50--154 dm_control/entities/manipulators/kinova/jaco_hand.py, lines 33--170 dm_control/manipulation/shared/robots.py, lines 25--53
Signatures	JacoArm(name=None) JacoHand(name=None, use_pinch_site_as_tcp=False) make_arm(obs_settings) -> JacoArm make_hand(obs_settings) -> JacoHand
Import	from dm_control.entities.manipulators import kinova from dm_control.manipulation.shared import robots

I/O Contract

Inputs

Function	Parameter	Type	Description
JacoArm	name	str or None	Optional prefix for MJCF model name attributes.
JacoHand	name	str or None	Optional prefix for MJCF model name attributes.
JacoHand	use_pinch_site_as_tcp	bool	If True, the pinch site is used as the tool center point instead of the grip site.
make_arm	obs_settings	ObservationSettings	Configures which arm observables are enabled and their processing options.
make_hand	obs_settings	ObservationSettings	Configures which hand observables are enabled and their processing options.
set_grasp	physics	mjcf.Physics	The physics instance to modify.
set_grasp	close_factors	float or Iterable[float]	Desired finger positions: 0 = fully open, 1 = fully closed.

Outputs

Function	Return Type	Description
JacoArm()	JacoArm	A composer entity with 6 joints, 6 velocity actuators, 6 torque sensors, and a wrist attachment site.
JacoHand()	JacoHand	A composer entity with 3 finger joints, 3 velocity actuators, grip/pinch sites, and finger/hand geom lists.
make_arm(obs_settings)	JacoArm	A JacoArm with manipulation-default observable options applied.
make_hand(obs_settings)	JacoHand	A JacoHand with use_pinch_site_as_tcp=True and manipulation-default observable options applied.

Usage Examples

from dm_control.manipulation.shared import robots
from dm_control.manipulation.shared import observations

# Create arm and hand with perfect-features observation settings.
obs_settings = observations.PERFECT_FEATURES
arm = robots.make_arm(obs_settings)
hand = robots.make_hand(obs_settings)

# Attach hand to arm at the wrist site.
arm.attach(hand)

# Inspect the arm's joints and actuators.
print('Arm joints:', [j.name for j in arm.joints])
# ['joint_1', 'joint_2', 'joint_3', 'joint_4', 'joint_5', 'joint_6']

print('Arm actuators:', [a.name for a in arm.actuators])
# ['joint_1', 'joint_2', 'joint_3', 'joint_4', 'joint_5', 'joint_6']

# Inspect the hand's finger actuators.
print('Hand actuators:', [a.name for a in hand.actuators])

from dm_control.entities.manipulators.kinova import jaco_hand

# Direct instantiation of JacoHand with pinch site as TCP.
hand = jaco_hand.JacoHand(use_pinch_site_as_tcp=True)
print('Tool center point:', hand.tool_center_point)
print('Number of finger joints:', len(hand.joints))
# 3

Related Pages

• Principle:Google_deepmind_Dm_control_Robot_Configuration