Geometric Fabrics

An NVIDIA reactive motion-generation framework that synthesizes obstacle avoidance and joint-limit behavior in real time with provable stability.

Proposed by Karl Van Wyk, Nathan Ratliff, and colleagues at NVIDIA Research (IEEE RA-L 2022), this is the successor to Riemannian Motion Policies (RMP). Rather than planning a whole trajectory in advance, it computes joint acceleration directly every control cycle: behaviors such as approaching a goal, avoiding obstacles, staying away from joint limits, and holding a posture are each written as a second-order differential equation in their own space, then combined into joint space through Jacobian mappings. Compared to RMP, it uses more general Finsler geometry, allowing behaviors to be shaped more flexibly while still guaranteeing stability. It's often used as a safe action space for reinforcement learning: DextrAH-G has the policy output only low-dimensional targets like a palm pose, leaving collision avoidance and joint constraints to the fabric. NVIDIA has open-sourced a GPU-parallel FABRICS library built on Warp on GitHub.

ExampleIn DextrAH-G, a Kuka arm plus an Allegro hand with 23 motors total has an RL policy output a palm target and low-dimensional finger commands, while the fabric converts these into joint commands at 60 Hz, automatically avoiding the table and self-collisions.

Also called

Fabrics

Related

Riemannian Motion Policies · DextrAH-G · Obstacle Avoidance · Joint Limits · Motion Planning · Operational Space Control

Sources

Van Wyk et al., Geometric Fabrics: Generalizing Classical Mechanics to Capture the Physics of Behavior (arXiv 2109.10443)
NVlabs/FABRICS (GitHub)
DextrAH-G: Pixels-to-Action Dexterous Arm-Hand Grasping with Geometric Fabrics

As of

2026-09