OPEN MODELS/WORLD MODELS/DREAMERV3-ROBOTICS
RESEARCHWORLD MODELS

DreamerV3-Robotics

DreamerV3 tuned specifically for robotic control tasks

WHAT DOES THIS ACTUALLY DO?

This is DreamerV3 adapted specifically for robotics tasks — with extra tuning for the kinds of movements and physics robots deal with. If you want to train a robot policy from scratch in simulation, this is a practical starting point.

PARAMETERS
~20M
RELEASED
2024-02
LICENSE
MIT
ARCHITECTURE
RSSM...
CAN I ACTUALLY RUN THIS?
5
Research
OUT OF 10

Community-maintained and less polished than the original DreamerV3. For researchers already familiar with model-based RL who want robotics-specific benchmarks.

DOWNLOAD WEIGHTS
View on GitHub
2GB disk space required
OVERVIEW

PyTorch implementation of DreamerV3 with robot-specific optimizations for manipulation and locomotion benchmarks.

TRAINED ON

Robotic manipulation and locomotion tasks (MuJoCo, Isaac)

ARCHITECTURE
RSSM World Model + Actor-Critic (PyTorch)
CAPABILITIES
Robot-specific tuning
PyTorch implementation
MuJoCo/Isaac Gym integration
Custom reward shaping
REAL ROBOTICS USE CASES
RESEARCH

Pre-train locomotion policies for legged robots

SIMULATION

Train manipulation policies with DreamerV3 in Isaac Gym

HARDWARE REQUIREMENTS
MIN VRAM
8GB
REC VRAM
16GB
MIN RAM
16GB
DISK SPACE
2GB
JETSON
Not Supported
CPU ONLY
GPU Required
TESTED GPUs
RTX 3080A100
DEPLOYMENT PATHS
Local PC
SUPPORTED
Docker
SUPPORTED
Cloud
SUPPORTED
Isaac Sim
SUPPORTED
ROS2
NOT SUPPORTED
Jetson
NOT SUPPORTED
RELATED PAPERS
Mastering Diverse Domains through World Models
RELATED ORGANIZATIONS
Google DeepMind
ORGANIZATIONS
Community (based on DreamerV3 by DeepMind)
DIFFICULTY
RESEARCH

Lab-grade infrastructure needed

COMPUTE4/10
SETUP3/10
BEGINNER-FRIENDLY3/10
EDGE DEPLOYMENT
Jetson Not Supported
CPU-only Not Available
Community →

Seeded and planned prompts — not live forum activity yet.