This library implements an inverse kinematics (IK) solver for dexterous multi-fingered hands. It takes as input Cartesian positions for the fingers of the hand and returns the joint angles needed to place the fingers at those desired positions.
The design of the IK solver is modular in that different solvers can be implemented by hot swapping the underlying controller. Specifically, the solver wraps an instance of dexterity.controllers, which is a library responsible for performing the mapping between Cartesian velocities and joint velocities. The current implementation relies on a simple Jacobian-based Damped Least Squares solver. In the future, we'd like to implement more sophisticated controllers, for example, ones that view IK under the lens of optimization.
from dexterity.inverse_kinematics import IKSolver
from dexterity.models import hands
hand = hands.ShadowHandSeriesE()
solver = IKSolver(hand)
# Suppose we are given a Cartesian position for each finger. The frame of reference is
# usually going to be at an `mj_site` on each fingertip.
target_positions = get_fingers_target_positions()
# An array of shape (num_fingers, 3).
qpos = solver.solve(
target_positions,
# How close we want to get to the target positions.
linear_tol=1e-3,
# How many IK attempts to perform. Each attempt randomizes the initial joint angles
# to arrive at a different solution.
num_attemps=5,
# Whether to exit at the first attempt that finds a solution within the tolerance.
stop_on_first_successful_attempt=True,
)
# If the solves fails, qpos will be None.
if qpos is not None:
hand.set_joint_angles(qpos)For a full-fledged example on how to use the solver, see examples/inverse_kinematics.py.
- The design philosophy of the IK solver, along with the controller it uses, is adapted from dm_robotic's ik_solver and C++ library.
- The Damped Least Squares controller is adapted from dm_control's inverse_kinematics.