Long-Range Gradients for Differentiable Contact Physics

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 220 papers26 conceptsfrozen 2026-10-04

In theory, differentiable physics simulators are a "cheat code" for robotic manipulation. If the entire simulation—from joint torques to the deformation of a piece of plasticine—is mathematically differentiable, we can use gradient-based optimization to find the perfect sequence of actions. Instead of trial-and-error reinforcement learning, we simply "backpropagate" through the physics to see how to move a hand to achieve a goal.

In practice, this approach often hits a wall: the zero-gradient problem. In standard physics models, if a robot hand is even a millimeter away from an object, the gradient of the goal (e.g., "pick up the cup") with respect to the hand's position is exactly zero. The simulator provides no "pull" to guide the hand toward the object, leaving the optimizer blind.

The Blind Spot of Hard Contacts

The fundamental issue is that physical contact is naturally discontinuous. As Marc Toussaint and colleagues note in their work on Logic-Geometric Programming, when a manipulator is not yet in contact, the objective function regarding the object's movement lacks any informative signal.

This "blind spot" persists across different materials. In PlasticineLab, Zhiao Huang and Hao Su demonstrate that while differentiable physics can solve soft-body tasks efficiently, the optimizer fails the moment contact is lost. Without a directional signal to re-establish contact, the system cannot plan multi-stage interactions. Similarly, in the DLO-Lab benchmark for cables and ropes, Junyi Cao and Chuang Gan highlight that gradients remain zero until the manipulated object actually touches the target, making it impossible to optimize trajectories from scratch.

Bridging the Gap with Long-Range Influence

To solve this, researchers are moving away from "hard" contact models toward "soft" or "long-range" gradients that provide a signal even at a distance.

One approach, pioneered by XIAOHAN YE, Kui Wu, and zherong pan, involves an Efficient Differentiable Contact Model with Long-range Influence. By ensuring the contact model is twice-differentiable and non-vanishing, they create a "force field" that guides the optimizer toward interactions even when objects start far apart.

Similarly, the DiffMJX framework, developed by Anselm Paulus and Andreas René Geist, introduces a mechanism called Contacts From Distance (CFD). As detailed in their recent work, CFD allows the simulator to produce accurate, informative gradients for collision-rich tasks by calculating potential forces before the collision occurs. This prevents the "dead zones" where an optimizer would otherwise stall.

Smoothing the Cost Landscape

Beyond modifying the physics engine itself, another strategy is to "smooth" the landscape the optimizer sees.

  • Diffusion Fields: Jan Peters and Georgia Chalvatzaki introduced SE(3)-DiffusionFields, which uses diffusion models to learn smooth cost functions. This turns a binary "grasp vs. no-grasp" signal into a continuous gradient field that pulls the robot toward successful configurations.
  • Hybrid Logic: Marc Toussaint suggests that instead of making everything differentiable, we should use first-order logic to search for discrete "modes" (like "pick up tool") and then use continuous optimization only once the contact sequence is decided.
  • Vision-Language Guidance: In DLO-Lab, the researchers use Vision-Language Models (VLMs) to provide a high-level plan, effectively giving the gradient-based optimizer a "warm start" near the first contact event.

Why It Works: From Points to Fields

The shift in this research area is a move from viewing physics as a series of discrete events (impacts) to viewing it as a continuous field of influence.

When Dylan Turpin and Miles Macklin use Grasp'D to synthesize multi-fingered grasps, the differentiability allows the system to discover "power grasps" with 4x denser contact areas than sampling-based methods. It works because the optimizer can "feel" the stability of the grasp improving with every tiny adjustment of the fingers—a signal that is only possible when the contact model provides a smooth, non-zero gradient across the entire search space.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 17 papers
  2. Version 1 · 2026-10-04 · 20 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_complementarity_aware_gradient_projectio_16d877a9,
  title = {complementarity-aware gradient projection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/complementarity-aware-gradient-projection/rdr_16d877a9c3b24db58212ea9fe6132272},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_long_range_gradients_for_differentiable__16d877a9_v1,
  title = {Long-Range Gradients for Differentiable Contact Physics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/long-range-gradients-for-differentiable-contact-physics/rdr_16d877a9c3b24db58212ea9fe6132272/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}