RBDReference

RBDReference is a CPU reference implementation for inspecting and validating rigid-body algorithms. It consumes the same parsed model as GRiD’s generator; it is not the batched GPU handle.

Run a reference calculation

After installing GRiD from its recursive source checkout, run this from the repository root:

import numpy as np
from URDFParser import URDFParser
from RBDReference import RBDReference

robot = URDFParser().parse("config/robot_assets/iiwa14.urdf")
if robot is None:
    raise ValueError("URDF parsing failed")
rbd = RBDReference(robot)
q = np.zeros(robot.get_num_pos())
qd = np.zeros(robot.get_num_vel())
qdd = np.zeros_like(qd)
tau, v, a, f = rbd.inverse_dynamics(q, qd, qdd, GRAVITY=-9.81)
M = rbd.crba(q)
Minv = rbd.minv(q, output_dense=True)
print(tau.shape, M.shape, Minv.shape)

This example is fixed-base and uses one state, not a batch. For quaternion models initialize a valid configuration (a zero quaternion is not a valid orientation) and use the model’s position/tangent coordinate maps. The reference’s crba takes q; a second positional argument is a normalization flag, not joint velocity.

Methods and validation

The reference includes inverse dynamics and its derivatives, CRBA, Minv, ABA, forward-dynamics derivatives, kinematics, centroidal quantities and plant/cost operations. The algorithm pages distinguish reference calls from the generated GPU handle calls.

For the maintained correctness workflows, see CUDA Validation And Performance Reporting and external/RBDReference/tests. Historical root-level scripts such as printReferenceValues.py are not part of this checkout’s supported workflow.

Dependencies

The pure-Python reference requires NumPy, SymPy and URDFParser. Optional Pinocchio-backed validation and second-order extensions have additional dependencies; follow the repository installation and validation instructions rather than treating NumPy alone as a complete oracle environment.