RobotGuide

Frontier · where the field is going

World models: six real directions, and how far a hype post is from them

World models are the main battleground of robot learning in 2025-2026: DreamerV3 made it into Nature, Meta's V-JEPA 2 achieved 65-80% pick-and-place in new environments with just 62 hours of real-robot data, and NVIDIA treats Cosmos as the "LLM of physical AI". This section starts from a "six major research directions" list circulating on Xiaohongshu — the directions are real, but the original post is nothing but stacked buzzwords with zero verifiable entities. Here each of the six directions is broken down into: the essence with the buzzwords stripped away, real paper anchors (with identifiers), the technical route the post never told you, the math background required, and entry-level experiments you can actually run.

Positioning: this is the fourth layer. /k/ is the toolkit, /tut/ is practice, /fund/ is algorithmic derivation, /math/ is mathematical principle — this layer covers "where the field is heading, and why". How to read: first check the relationship diagram on the index page to pick a route, then choose one direction and enter its detail page; every experiment on the detail pages has actually-run numbers.

The six directions at a glance

1 Continual Reinforcement Learning: Learn the New Without Forgetting the OldWhen an agent learns task A and then task B, the shared parameters get dragged away by the new data and A's capability collapses — catastrophic forgetting. How to avoid forgetting on the world model track. 18 min3 paper anchors 2 Open-Environment Adaptation: Pre-train a Worldview First, Patch Tasks Online LaterFirst pre-train a "general understanding of the world" from large amounts of offline data (no reward, non-expert); after deployment, fine-tune to specific tasks with a little online interaction; when encountering unseen environments, temporarily piece together multiple domain-expert models. 20 min4 paper anchors 3 Uncertainty Quantification: Let the Model Know What It Doesn't KnowWhen the model is wrong it must know it is wrong. Turn "I'm not sure" into a computable number: avoid high-uncertainty regions when planning, and actively go to high-uncertainty regions when exploring. 16 min4 paper anchors 4 Efficient Inference: Planning Pays Compute for Returndecision-time planning must try hundreds of action sequences in latent space at every step (TD-MPC2 defaults to 512 candidates × 6 iterations); real-robot control frequency cannot bear it; the longer the rollout, the greater the compounding of one-step errors. How to trade less search for the same decision quality. 18 min3 paper anchors 5 Hierarchical Control: Macro-actions Save 480× the Imagined ComputeOn long-horizon tasks, let a slow and abstract high level emit subgoals/macro-actions and a fast and concrete low level emit raw actions, with the two layers aligned in latent space. The origin is Sutton 1999's Options framework; the world model version = the high level plans macro-action sequences in imagination. 18 min4 paper anchors 6 Sim-to-Real: Pixel Gap Is the Appearance, Physics Gap Is the EssenceWhat is learned in simulation fails on the real robot — the two worlds' conditional distributions are not the same one. The world model route's distinctive answer: rather than fixing the simulator, let the model learn "what physics looks like" by itself from real videos. 18 min4 paper anchors

Relationship map: what is the foundation of what

                 ┌─ Direction 3  Uncertainty quantification (component layer, called by 1/2/4/6)
                 │
World-model      ┼─ Direction 1  Continual RL  ←──  Direction 2  Open-environment adaptation
backbone         │                                (reusable priors first, then continual learning)
(RSSM/Dreamer)   ├─ Direction 4  Efficient inference ←≈→ Direction 5  Hierarchical control
                 │                        (two sides of one coin: temporal abstraction =
                 │                         fewer decision points = compute saved)
                 └─ Direction 6  Sim-to-real (the acceptance gate the backbone and every
                                  component must pass in the end)

Suggested learning order

  1. Run one Dreamer-family practice end to end, with the whole RL derivations track underneath — stand up the backbone before picking a direction.
  2. Start Direction 3 (Experiment A, about two days) — the cheapest first real run, with probability and conditional expectation as the base.
  3. Close-read Direction 1/2 papers: arXiv 2507.09177 → DINO-WM → arXiv 2502.19544.
  4. Hands-on for Directions 4/5 (Experiments C+D) — they share one paper pool (HWM/FOCUS); one step from the backbone to a runnable experiment.
  5. Finish with Direction 6 — the acceptance gate everything must pass; the labs are the place to build the reflexes.

The gap from the original post

All six directions are real; as information delivery, all six fail: the original post has no papers, no method names, no benchmarks. The least frontier is Direction 3 (PILCO was doing it in 2011); the most mislabeled are Direction 1 (parameter-efficient fine-tuning is not the mainstream fix) and Direction 6 (domain adversarial stalled around 2018); the best-earned hype is Direction 2 and Direction 5 (the two lines with the densest 2025-2026 paper output). If you only pick one to go deep: read Direction 4+5 together.