← Paper Digest

Alloy: A Multi-Material Physics Engine for Robot Learning

Pranay Gupta, Deepanshu Rohilla

November 2026

Highlight reel of Alloy scenes, from pouring water to quadrupeds and humanoids.

TL;DR Alloy is an open-source engine that puts rigid robots, cloth, cables, liquids, granular material and soft bodies into one scene. It routes each object to a suitable solver and couples the solvers inside every control step, across many parallel worlds on one GPU. Coding agents that had never seen it trained policies for quadruped walking (97% held-out success), cable lifting (95.7%) and pouring (92%) on a single L4 GPU.

What they do

Simulators for robot learning at scale are strongest on rigid bodies, and adding cloth, cables or liquids usually means a second engine plus hand-written glue. Alloy puts all of these materials in one scene and runs it in many parallel worlds on one GPU, or on a CPU. It is built on Newton, MuJoCo Warp and Warp, and is designed to be driven by coding agents.

How it works

  • Routing. Each object is a shape plus a material card (density, stiffness, friction). The router assigns it a solver and records why: MuJoCo Warp for rigid bodies and robots, Newton’s VBD for cloth, soft volumes and rods, implicit MPM for sand and pastes, Alloy’s SPH for liquids.
  • Coupling. Solvers exchange poses and contact forces through proxy bodies inside every control step. The step size comes from the interface’s effective mass and contact stiffness. If a world’s interface work looks wrong, that control period is rerun with a finer step.
  • Health grades. Every world is graded OK, DEGRADED or INVALID after each exchange. Invalid worlds are reset and frozen, separating physics failures from policy failures.
  • Agent interface. The docs ship inside the package, every error names its fix, and the engine is exposed over the Model Context Protocol.

Results

  • Coding agents on one NVIDIA L4, reading only the shipped docs, trained Go2 walking to 97% held-out success in 15.5 min (97-98% under 150 N pushes). Cable lifting reached 95.7% in about 30 min, pouring 92% in 38.5 min, and behavior-cloned pick-and-place 256/256.
  • From 1 to 64 worlds, throughput grows about 39x for a Franka scene, 20x for a cloth sheet and 13x for a 200 ml pour.
  • All ten closed-form checks (e.g. cantilever bending, catenary sag, hydrostatic pressure) fall within tolerances set in advance. A ten-block tower had no errors in any of 1,024 worlds.

Limitations

Alloy is an early 0.x release whose interfaces may change. The policy results come from a pre-release, one training run each. Everything is in simulation, with no transfer to real robots.

Videos

Two grippers fold a towel.
Grippers pass a thread through the eye of a needle.

Figures

From Figure 1: A Franka arm pours water from a bottle into a glass; the water is simulated with Alloy's liquid solver...
From Figure 1: A Franka arm pours water from a bottle into a glass; the water is simulated with Alloy's liquid solver, coupled to the robot's rigid-body contact.
From Figure 1: 1,024 parallel worlds of one robot task in a single frame.
From Figure 1: 1,024 parallel worlds of one robot task in a single frame.
Abstract

Robots handle much more than rigid objects: towels, cables, water, sand. The simulators used to train robots at scale are strongest on rigid bodies, and adding another material usually means a second engine and hand-written glue between the two. We present Alloy, an open-source physics engine for robot learning in which rigid robots, cloth, cables, liquids, granular material and soft bodies share one scene. Each object is described once; Alloy assigns it to a solver suited to its material, couples the solvers inside every control step, and runs the scene in many parallel worlds on one GPU, or on a CPU. Every world carries a health grade, to help tell a failure of the physics from a failure of the policy. Alloy is built on Newton, MuJoCo Warp and Warp, and it is written to be driven by coding agents: its documentation ships inside the package, every error names its fix, and the engine is exposed as tools over the Model Context Protocol. In a cold test, agents that had never seen Alloy installed a pre-release from the package index and trained policies for walking, cable handling and pouring on a single GPU. We also report how throughput grows with the number of worlds, depth-camera throughput, agreement with closed-form references, and a reliability test over a thousand worlds.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.