lucidrains/rigidformer

★ 13PythonAudience · researcherComplexity · 4/5ActiveSetup · hard

Mindmap

mindmap
  root((rigidformer))
    Inputs
      Rigid body states
      arXiv paper
    Outputs
      Predicted next state
      Python modules
    Use Cases
      Study attention for physics
      Re-implement the paper
      Run rigid dynamics models
    Tech Stack
      Python
      PyTorch
      Transformers
      Attention

mindmap root((rigidformer)) Inputs Rigid body states arXiv paper Outputs Predicted next state Python modules Use Cases Study attention for physics Re-implement the paper Run rigid dynamics models Tech Stack Python PyTorch Transformers Attention

Things people build with this

USE CASE 1

Track an open re-implementation of the RigidFormer paper as it lands

USE CASE 2

Study how attention layers can be applied to rigid body dynamics prediction

USE CASE 3

Read the Python modules alongside the arXiv paper to learn the architecture

USE CASE 4

Borrow building blocks for your own physics-aware Transformer experiments

Tech stack

PythonPyTorch

Getting it running

Difficulty · hard Time to first run · 1day+

README has no install or usage instructions; you have to read the arXiv paper and the source to figure out what is implemented.

In plain English

Rigidformer is an early, work-in-progress Python implementation of a research paper called RigidFormer: Learning Rigid Dynamics using Transformers. The paper comes out of MIT and Meta and is hosted on arXiv. The point of the underlying research, judging from the title and topics tagged on the repository, is to predict how solid (rigid) objects move and collide, using a Transformer neural network of the kind first popularised in natural language processing. The README itself is very sparse. It contains a single small figure (fig2.png), a one-line title flagged as wip (work in progress), a single sentence pointing at the arXiv paper and naming the institutions behind it, and a BibTeX citation block listing the authors (Zhiyang Dou, Minghao Guo, Haixu Wu, Doug Roble, Tuur Stuyck, and Wojciech Matusik) and the publication year 2026. There is no install command, no usage example, no description of the model architecture, no training instructions, and no link to weights or data in the README. This matches the wider pattern of the author's published work (lucidrains is well known for open-source PyTorch re-implementations of recent papers), but a reader who lands here today will not learn from the README alone what classes the code exposes or how to run it. The honest reading is that this is an in-progress code drop that points readers at the original arXiv paper for any explanation of what the model does or how it is trained. The repository metadata adds a little more context that the README does not spell out. The primary language is Python, which is consistent with the author's usual PyTorch implementations. The topics list on the repository page tags the project as artificial intelligence, attention mechanisms, deep learning, and rigid body dynamics, suggesting that the model uses attention layers to learn how to predict the next state of a physical scene made of solid objects. A reader who wants to understand or use the code today will almost certainly have to read the linked arXiv paper first, then look directly at the Python source in the repository to see which modules and classes have been implemented so far.

Copy-paste prompts

Prompt 1

Read the Python source in lucidrains/rigidformer and map each class back to a section of the arXiv paper

Prompt 2

Sketch a minimal training loop for rigidformer on a small rigid body dataset like Kubric

Prompt 3

Compare the attention pattern in rigidformer with a standard sequence Transformer block

Prompt 4

Outline a usage README for rigidformer based on the public modules currently in the repo

Prompt 5

Stub out a unit test that runs one forward pass through the main rigidformer model class

Open on GitHub → Explain another repo

Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.