latent-dynamics

INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models.

A modular PyTorch library designed for learning, training, and deploying world models across various environments.

Curated papers, code, datasets, and benchmarks for medical world models in imaging, EHR trajectories, treatment planning, surgical AI, robotics, and virtual-cell simulation.

Learning shared latent dynamics across multiple asynchronous, non-overlapping neural recording sessions (ICLR 2025).

Action-conditioned JEPA world model for two-room navigation in 89K parameters. VICReg plus an inverse-dynamics auxiliary loss for collapse resistance; 4.40 MSE on normal probes.

Evidence-bound world-model integration: motion baseline and pinned neural adapters.

Demo implementations of JEPA World Models to support research

CARDIOKOOP - Control-aware Koopman deep learning framework for real-time hemodynamic forecasting and cardiovascular digital twin applications.

Compact, action-conditioned latent world models for planning and control.

CARDIOKOOP - Control-aware Koopman deep learning framework for real-time hemodynamic forecasting and cardiovascular digital twin applications.

V-JEPA for Gray-Scott dynamics. Initial work produced during the 24-hour Hack the World(s) hackathon. 1st place 🏆

Minimal reproducible framework for world-model rollout: measures how many steps a learned world model stays trustworthy (reliable horizon) before its own predictions degrade it.

Conservative Lapse-Action Planning (CLAP): a variational access-and-dwell framework for safe AI optimization. Surf the conservative-lapse field to the best safe basin and dwell there — theorem-backed, pip-installable, with differentiable PyTorch training adapters.

World model on a Boolean hypercube: frozen reservoir encoders and a locally connected predictor that learns how the latent code moves under an action. C++20, no dependencies.

A PET MOVIE for Interpretable Synthesis of Late-Frame [11C]-PiB PET Images from Early-Frame Counterparts

Analysis of premotor cortex signals in macaques during a countermanding task, using a PyTorch Deep Markov Model to generate and simulate neural activity in a low-dimensional latent space.

Decoupled autoencoder framework: bounded + unbounded latent carriers with two-stage frozen training. Lorenz-63 chaotic system benchmark.

RSSM-based world model for latent market dynamics, regime detection, and risk management. DreamerV2-style architecture.

Emergentia is a neural-symbolic discovery engine that extracts parsimonious physical laws from noisy particle trajectory data. It combines deep learning to model complex forces with symbolic regression to rediscover human-readable, mathematically interpretable equations of motion.

Latent dynamics model plus MPPI planning for cooperative multi-agent RL in Overcooked-AI. I built the Q-value network and training stack (merged upstream).