dreamerv3

[ICLR 2025 Oral] PyTorch code for the paper "Open-World Reinforcement Learning over Long Short-Term Imagination"

Modular DRL framework for autonomous robot navigation in ROS2. Plug-and-play RL backends (Stable-Baselines3, DreamerV3), composable reward functions, observation spaces & neural architectures - built for research and deployment.

Flax Implementation of DreamerV3 on Crafter

PyTorch implementation of DreamerV3 from "Mastering Diverse Domains with World Models"

Autonomous FPV drone control with pure vision input

DreamerX is an implementation of model-based DreamerV3 with minor optimizations and novel training adjustments. It is designed to be flexible and user-friendly, allowing researchers and practitioners to easily interchange components and environments.

DreamerV3 World Model RL from Scratch — Educational implementation of model-based reinforcement learning

Hierarchical learning by dreaming for empowering control with latent skills in imagination

The implementation of pytorch-based DreamerV3 for Meta-world simulator.

[ICLR 2025 Oral] PyTorch code for the paper "Open-World Reinforcement Learning over Long Short-Term Imagination"

The World Model Remembers, the Actor Forgets: Dream Rehearsal for Continual Model-Based RL — code, pre-registration trail, and run data

Spatial memory for Physical AI. Robots, drones & world models remember what they saw, where and when, and notice what's new. 4D (x,y,z,t) vector memory · novelty detection · MCP server.

Hardened FH6 self-driving research stack: screen capture, Forza Data Out telemetry, DreamerV3, AutoDrive recovery demos, and route-aware stress-tested training.

A template for deploying DreamerV3 with Ray RLlib, compatible with Gym and custom environments.

Experimental Model-Based RL agent for Pokémon Red. Implements continuous (VAE+GRU) and discrete Recurrent State-Space Models (RSSM) to learn environment dynamics. Features ablation studies on latent imagination drift, scheduled sampling, and MPC planning, inspired by Dreamer and DIAMOND architectures.

Uncertainty-gated imagination learning and failure-safe training for DreamerV3.

🌍 Investigating the understanding of spatio-temporal information in World Models | Research Project in World Models 2024 by The University of Tokyo

双语技术博客 · 世界模型 / RSSM / DreamerV3 / VLA / Sim-to-Real

Reinforcement Learning : Autonomous parallel parking task. implementing SAC and DreamerV3's World Model on Highway-env

Extending DreamerV3 for robotic application and better Morphologieawareness by implementing a FFKSM