Feynman Letter: Compositional Diffusion with Guided Search for Long-Horizon Robot Planning

Forum topic · 小凯 · 2026-05-03

Summary

This post reviews the robotics control paper 'Compositional Diffusion with Guided Search' (2026.05) and its approach to long-horizon planning in embodied AI. The author explains why current reinforcement-learning robots fail at multi-step tasks like storing an apple in the fridge when obstacles require dozens of sequential actions—a problem framed as 'causal chain breakage from Markovian myopia.' The paper's solution combines three ideas: generating state trajectories as images using diffusion models (similar to how video models synthesize motion), compositionality (mathematically combining probability distributions of subtasks without requiring exact demonstration videos), and guided search, which constrains physically infeasible generated trajectories toward realistic collision boundaries. The author frames long-term planning as backward reasoning from a clear final physical state, arguing that true household robotics requires a 'spacetime simulator' inside the robot's silicon brain. The takeaway: for problems with long feedback cycles, build a global state diffusion mechanism rather than optimizing greedily step by step, since local optima without a visible endpoint may accelerate failure.

Feynman Letter: Do You Want a Robot That 'Improvises Step by Step', or One That 'Plays a Movie in Its Head'?

A review of Compositional Diffusion with Guided Search (2026.05), a robotics control paper. After reading it, I feel that in teaching robots to do housework, humanity has finally given up on rote memorization and started teaching them to dream.

To understand why today's robots freeze up in even a slightly complicated kitchen, let's talk about cooking.

1. Current state: the nearsighted sweeper-monk

Today's robots (e.g., early reinforcement learning models) are like an extremely short-sighted sweeper-monk.

Pain point: Ask it to "put the apple in the fridge." If the apple is right in front of it, it does a fine job. But if the apple is in a cabinet and the fridge door is closed, it instantly gets confused. Because its brain (the algorithm) can only predict the next one second of motion. Facing long-horizon planning that needs dozens of steps (open the cabinet -> take out the apple -> open the fridge -> put in the apple), it gets lost at step one. This is the causal chain breakage caused by Markovian myopia.

2. Compositional diffusion with guided search: a strategist with a built-in "4D sandbox"

The paper's idea is brutally simple and elegant: since you can't compute every future step, let the diffusion model directly generate the 'image of the future'!

It achieves a two-layer leap for embodied intelligence along the time axis:

Physical imagery (diffusion generation of state trajectories): It treats the robot's action sequence as a "picture" to be generated (Diffusion Models). Like Sora generating video, the model directly "diffuses" a whole coherent action sequence—from opening the cabinet to closing the fridge—in latent space.
Compositionality: You don't need to show it cooking videos that are exactly identical to the target task. It can mathematically superpose the probability distribution of "open the cabinet" with that of "open the fridge".
Guided search: This is its obstacle-avoidance radar. When the diffusion model generates a wonderfully dreamy trajectory that would physically smash into a wall, the guided search algorithm acts like a strict physics teacher, forcibly bending that imaginary trajectory back into the real boundaries of physical collision.

3. A Feynman-style judgment: planning is "backward inference along the arrow of time"

So-called "long-term planning" was never about building an endless if-else tree in your head.

It is this: you first see, with extreme clarity deep in your mind, that ultimate physical state (the apple is already in the fridge), then, following the causality of time, collapse backward step by step into the action of the present.

Compositional diffusion control tells us: what makes a robot smart is not just more motors, but whether its silicon brain can hold a "multi-dimensional spacetime simulator" of its own.

Only when a robot can, before moving its first finger, have already "played back" the entire housework flow like a movie in its subconscious will the era of general-purpose household robots truly arrive.

Takeaway

When solving any complex problem with an ultra-long feedback cycle, stop grinding on "what's the next step."

Go build your "global state diffuser".

If you cannot see the endpoint from the start, then every locally optimal move you make along the way may just be accelerating your rush toward the cliff.

*Source hashtags: EmbodiedAI, Robotics, DiffusionModels, LongHorizonPlanning, ReinforcementLearning, FeynmanLearning.*