Closed kinematic loops—where links form cycles rather than simple trees—are common in many robotic mechanisms, but they are notoriously awkward to simulate. Most physics engines sidestep the issue by approximating the robot as a kinematic tree or by adding explicit loop-closure constraints. A new paper from researchers at NVIDIA and Disney Research introduces Kamino, a GPU-based solver that handles these couplings directly, without such workarounds.
Kamino is implemented in Python using NVIDIA Warp and integrated into the Newton framework. At its core is a constrained optimization algorithm based on ADMM, which computes constraint forces by solving the rigid multi-body forward dynamics as a nonlinear complementarity problem. This approach preserves contact dynamics without resorting to simplified or convexified models, enabling high-fidelity simulation of highly coupled systems.
The authors demonstrate the method by training a reinforcement-learning policy for a bipedal robot with six nested kinematic loops. Kamino simulated 4,096 parallel environments on a single GPU and produced a feasible walking policy. The solver also supports heterogeneous worlds, meaning structurally different robots can be batched together in the same simulation, which could be useful for large-scale data-driven robotics research.