DreamerV3 on WorldSense Tech Blog

What Does a World Model Actually Do in a Robot? From Perception to Action

Wed, 26 Aug 2026 00:00:00 +0000

Dreamer Series · Part 2

Part 1 broke down Dreamer’s overall architecture. This article zooms out one step further: where exactly does a world model sit in a robotic system? What happens between sensor data and the final action? Rather than discussing how to train a general-purpose robot world model, this article focuses on its functional position within robotic systems.

1. A Fundamental Difference: Robot World Models Are Nothing Like Language Models

Over the past few years, the word “model” has been used repeatedly across AI. Language models predict the next token, vision models predict the next frame, autonomous driving models predict the behavior of traffic participants. They all do some form of “prediction,” but the objects and constraints differ enormously.

Understanding Dreamer: How World Models Learn to Imagine

Tue, 25 Aug 2026 00:00:00 +0000

Dreamer Series · Part 1

This article covers Dreamer’s overall architecture at a conceptual level. If you’ve already read the RSSM Code Walkthrough Series, this article will help you connect the scattered code details into a coherent architectural understanding.

Dreamer’s most important idea is not “training a model that generates future frames,” but rather training a latent world model sufficient to support decision-making, then letting policies learn inside that internal world.

Understanding RSSM Through Code (6): Default Config, Four Formulas, and the Code↔Math↔Semantics Map

Mon, 24 Aug 2026 00:00:00 +0000

Understanding RSSM Through Code · 第 6 篇 / 共 6 篇

Series contents (you are on part 6, bolded; prev/next at the bottom):

(1) Where RSSM Sits & the Stochastic State

(2) Prior/Posterior, Straight-Through & unimix

(3) _core(), deter=8192 & Block GRU

(4) KL Balancing, Free Nats & Final KL

(5) Imagine, Observe vs. Imagine & Reset 6. (6) Default Config, Four Formulas & the Map

25. Default Config: Split into “RSSM Architecture” and “World-Model Training”

To avoid mixing “RSSM architecture parameters” with “world-model / agent training config,” the configuration is split into two tables below.

Understanding RSSM Through Code (5): Imagine, Observe vs. Imagine, Sequence Training, and Reset

Sun, 23 Aug 2026 00:00:00 +0000

Understanding RSSM Through Code · 第 5 篇 / 共 6 篇

Series contents (you are on part 5, bolded; prev/next at the bottom):

(1) Where RSSM Sits & the Stochastic State

(2) Prior/Posterior, Straight-Through & unimix

(3) _core(), deter=8192 & Block GRU

(4) KL Balancing, Free Nats & Final KL 5. (5) Imagine, Observe vs. Imagine & Reset

(6) Default Config, Four Formulas & the Map

20. Imagine: How Does RSSM Run Without Observations?

This is the most elegant part of RSSM.

Understanding RSSM Through Code (4): KL Balancing, Free Nats, and the Final KL Combination

Sat, 22 Aug 2026 00:00:00 +0000

Understanding RSSM Through Code · 第 4 篇 / 共 6 篇

Series contents (you are on part 4, bolded; prev/next at the bottom):

(1) Where RSSM Sits & the Stochastic State

(2) Prior/Posterior, Straight-Through & unimix

(3) _core(), deter=8192 & Block GRU 4. (4) KL Balancing, Free Nats & Final KL

(5) Imagine, Observe vs. Imagine & Reset

(6) Default Config, Four Formulas & the Map

14. Complete Observe Phase Data Flow (Observe / Imagine in One Framework)

We can now draw the RSSM state-transition framework and put Observe and Imagine inside the same _core() transition framework:

Understanding RSSM Through Code (3): Deterministic Transition _core(), deter=8192, and Block GRU

Fri, 21 Aug 2026 00:00:00 +0000

Understanding RSSM Through Code · 第 3 篇 / 共 6 篇

Series contents (you are on part 3, bolded; prev/next at the bottom):

(1) Where RSSM Sits & the Stochastic State

(2) Prior/Posterior, Straight-Through & unimix 3. (3) _core(), deter=8192 & Block GRU

(4) KL Balancing, Free Nats & Final KL

(5) Imagine, Observe vs. Imagine & Reset

(6) Default Config, Four Formulas & the Map

9. The Real Deterministic Transition: `_core()`

This is the most worthwhile part of rssm.py to read.

Understanding RSSM Through Code (2): Prior/Posterior, Straight-Through Sampling, and unimix

Thu, 20 Aug 2026 00:00:00 +0000

Understanding RSSM Through Code · 第 2 篇 / 共 6 篇

Series contents (you are on part 2, bolded; prev/next at the bottom):

(1) Where RSSM Sits & the Stochastic State 2. (2) Prior/Posterior, Straight-Through & unimix

(3) _core(), deter=8192 & Block GRU

(4) KL Balancing, Free Nats & Final KL

(5) Imagine, Observe vs. Imagine & Reset

(6) Default Config, Four Formulas & the Map

4. Translating the Source Code into Mathematical Formulas (Watch the Time Index)

DreamerV3’s recurrence relation:

Understanding RSSM Through Code (1): Where RSSM Sits in DreamerV3 and the Stochastic State

Wed, 19 Aug 2026 00:00:00 +0000

Understanding RSSM Through Code · 第 1 篇 / 共 6 篇

Series contents (you are on part 1, bolded; prev/next at the bottom): 1. (1) Where RSSM Sits & the Stochastic State 2. (2) Prior/Posterior, Straight-Through & unimix 3. (3) _core(), deter=8192 & Block GRU 4. (4) KL Balancing, Free Nats & Final KL 5. (5) Imagine, Observe vs. Imagine & Reset 6. (6) Default Config, Four Formulas & the Map

Source Code Reading Guide

DreamerV3 GPU Infrastructure: Cloud vs Self-Built Cost Analysis

Mon, 17 Aug 2026 00:00:00 +0000

In previous articles, we covered DreamerV3 environment setup and training tips. But there’s a more fundamental question that often gets overlooked: what hardware are you running on?

This question seems simple, but it directly affects both your research pace and your wallet. I spent a month on AutoDL cloud GPUs, then a week planning a self-built workstation — and even received a ¥36,000 turnkey quote from a vendor. This article shares the entire process and cost breakdown, in the hope of giving anyone wrestling with the “rent vs. buy” decision a useful reference.

Isaac Lab: From DreamerV3 to Industrial-Scale Robot RL Training

Fri, 14 Aug 2026 00:00:00 +0000

Over the past week, we’ve gone deep on the MuJoCo + DreamerV3 pipeline — from environment setup and visual-input training, to training tricks and the evolution of world model architectures.

Today, let’s shift perspective and look at another tech stack: NVIDIA’s Isaac Lab.

If MuJoCo emphasizes lightweight, flexible dynamics research suited for rapid prototyping and algorithm exploration, then Isaac Lab emphasizes GPU-accelerated, large-scale robot training and sim-to-real pipelines. The two are not mutually exclusive — many research teams use MuJoCo for algorithm validation and Isaac Lab for large-scale training simultaneously.

When World Models Meet Transformers: From RSSM to Large-Scale Sequence Modeling

Wed, 12 Aug 2026 00:00:00 +0000

In previous articles, we covered the RSSM architecture and training techniques in DreamerV3 in depth. RSSM is a classic design in reinforcement learning world models, but if you follow recent research, you’ll notice a clear trend: world models are becoming Transformer-based.

From Google’s UniSim to Wayve’s GAIA-1, from NVIDIA’s Cosmos to solutions from domestic embodied AI teams, the Transformer is emerging as a key technical approach for large-scale world models.

DreamerV3 Training Tips: Lessons from Real-World Debugging

Tue, 11 Aug 2026 00:00:00 +0000

In the previous article, we walked through four representation approaches for world models. Today, we shift back to the practical side of DreamerV3 and talk about the pitfalls and tricks you encounter during training. This article is based on local experiments using DreamerV3 commit e3f02248, JAX + Haiku, and MuJoCo + DM Control. Parameter names and configurations may differ across versions.

DreamerV3 is currently one of the most open-source and mature world model implementations available. But if you’ve actually trained it, you know the process is far from easy — environment setup, hyperparameter tuning, training instability, slow convergence… the list of gotchas goes on.

Building a World Model Lab from Scratch: A MuJoCo + DreamerV3 Practical Guide

Wed, 05 Aug 2026 00:00:00 +0000

I’ve written several theoretical articles on world models so far — from the mathematics of RSSM to Sim-to-Real transfer, to the comparison between VLAs and world models. A reader asked: “I get the theory, but how do I actually run something?”

Today’s article answers that question. I’ll walk you step by step through setting up a complete world model experimentation environment — from installation to training to visualization. Once you’ve gotten through it, you can build your own experiments on this foundation.

Is World Model a Good Research Direction? An Engineer's Honest Assessment

Mon, 03 Aug 2026 00:00:00 +0000

As an engineer who has been working in this field for over half a year, here are my thoughts.

Let me start with the conclusion: it is a good direction, but not everyone should jump in right now.

Why It’s a Good Direction

World models address a very fundamental problem: enabling AI not just to “see” the world, but to “understand” it.

Large language models have already demonstrated that when a model is large enough and the data is sufficient, strong capabilities can emerge. But language models understand the world of text, not the physical world. For robots to truly operate in real-world environments, they need to understand physical laws — gravity, friction, collisions, causality. These things cannot be learned from text data alone.

What Is a Robot World Model? An Engineer's Deep Dive

Sat, 01 Aug 2026 00:00:00 +0000

If you’ve been following the latest developments in AI, you may have noticed a trend: from ChatGPT to Sora, from AlphaFold to robotic manipulation, AI is moving from “understanding language” to “understanding the world.” At the heart of this transition lies an increasingly central concept — the World Model.

In today’s post, I want to discuss, from an engineer’s perspective, what a world model is, why it matters so much for robotics, and what DreamerV3 — currently one of the most representative approaches — actually does.