MESSIER LABS

Research

Recovering signal from noise.

A catalogue of what we work on. Each area is a step toward encoders that turn raw, high-dimensional observation into compact latent state — the representation a world model needs to predict, plan, and generalize.

M1Representation

State from raw observation.

A world model is only as good as what it encodes. We learn compact latent states directly from raw, high-dimensional observations — pixels, sensors, streams — without handcrafted features deciding in advance what matters.

  • 1.1Encoders that compress noisy, high-dimensional input into a small, structured latent state.
  • 1.2Representations learned end-to-end from raw data rather than engineered pipelines.
  • 1.3Robustness to distractors, occlusion, and sensory noise that break brittle feature extractors.
M2Representation

Dynamics-aware encoders.

Reconstruction wastes capacity on everything visible; prediction spends it on what changes. We shape encoders by their ability to predict the future, so the latent state keeps only what is controllable and consequential.

  • 2.1Self-supervised objectives that reward predictive, not pixel-perfect, representations.
  • 2.2Latent spaces that separate what an agent can control from what it merely observes.
  • 2.3Encoders that discard irrelevant detail instead of memorizing it.
M3Dynamics

Rolling the world forward.

On top of a good encoder, we learn latent dynamics that stay stable over long horizons — a model that can imagine many steps ahead without drifting into noise.

  • 3.1Latent transition models that remain accurate far beyond a single step.
  • 3.2Long-horizon rollouts that stay physically and causally consistent.
  • 3.3Uncertainty that is represented, not hidden, as predictions run forward.
M4Control

Planning in latent space.

A world model earns its keep when an agent can plan inside it. We use learned encoders and dynamics for control — searching over imagined futures instead of expensive real-world trials.

  • 4.1Planning and policy learning carried out entirely in the learned latent space.
  • 4.2Sample-efficient control that transfers across tasks and embodiments.
  • 4.3Behavior grounded in a model of the world, not memorized trajectories.

Team

Two researchers working at the intersection of representation learning, world models, and generative modeling.

BL

Bruno Lima

Machine learning · generative models

Senior at Duke University studying Computer Science and Mathematics. Previously worked across big tech, quant trading, and AI startups — including Meta, Citadel Securities, and Metis (acquired by DoorDash). A medalist in the international physics and astronomy olympiads representing Brazil.

bruno.lima@duke.edu
WC

William Chen

Diffusion models · generative research

Senior at Duke University studying Computer Science and Mathematics. Two years of diffusion research at Duke, with two first-author NeurIPS papers under review. Teaching assistant for Algorithms.

william.chen992@duke.edu

Status

Early research. Papers and open-source releases forthcoming.

Get in touch
Built with v0