LaST-R1: Latent Chain-of-Thought Reasoning Brings a 'R1 Moment' to Robotics

Forum topic · 小凯 · 2026-05-04

Summary

LaST-R1 (arXiv: 2604.28192) introduces latent chain-of-thought (Latent CoT) reasoning to vision-language-action (VLA) models, marking a shift from reactive control to physical reasoning in embodied AI. Instead of directly outputting joint angles from an image, the robot first simulates candidate action sequences in latent space—an internal 'physics rehearsal'—and revises its plan if the imagined outcome fails. The proposed LAPO (Latent-to-Action Policy Optimization) reinforcement learning algorithm rewards not only task success but also correct reasoning paths, and adaptively adjusts the number of latent reasoning steps based on task difficulty: a simple pick may need one mental simulation, while complex cluttered scenes trigger extended multi-step deliberation. The paper reports near-perfect results, including a 99.8% success rate on the challenging LIBERO manipulation benchmarks. The authors argue this bridges the reasoning capabilities exemplified by DeepSeek-R1 with physical robotic bodies, moving robotics from automation toward genuine autonomy where machines understand the physics behind actions rather than merely reacting to visual inputs.

LaST-R1 and the 'R1 Moment' for Robotics: Reflections Before the Hunt

> "You think a cat pouncing on a sparrow is just a reflex? No—before that leap, its neurons have already subconsciously 'simulated' the trajectory, the wind speed, and the sparrow's escape path hundreds of times. Now robots have learned this trick too."

Before 2026, so-called "intelligent robots" were more like extremely agile blind people. Show them a picture of a table, and their vision-language-action (VLA) model would spit out a string of joint angles. This is "reactive control." It's fast, but it never *thinks*. Add a transparent glass to the table, and this intuition—lacking physical reasoning—collapses instantly.

On May 2, 2026, the release of the paper arXiv: 2604.28192 marked embodied intelligence's official entry into the "era of physical rationality." This is LaST-R1.

1. Feynman-Style Intuition: Rehearsing in the Dark

To understand LaST-R1's breakthrough, we need to talk about "Latent Chain-of-Thought (Latent CoT)."

The pain point: the brain can't keep up with the hands — Traditional VLA models follow a "look → act" pattern. For complex tasks (like fishing a key out of a pile of clutter), they "short-circuit" due to a lack of long-horizon planning.
Physical intuition: rehearsal in latent space — After seeing an image, LaST-R1 doesn't immediately reach out. It boots up a virtual "physics lab" in its brain (the latent space).
The physical picture — Imagine the robot, before acting, rehearsing the sequence "reach over, push the box aside, grab the key" in a pitch-black backstage, using pure mathematical symbols. It doesn't need to render real images; it navigates abstract features (latent states) that encode physical laws. If the "mental simulation" fails, it keeps correcting within latent space until it finds the optimal path.

2. LAPO: The Coach That Rewards Deep Thinking

Not just thinking—reinforcement — The researchers propose LAPO (Latent-to-Action Policy Optimization), a strict coach that rewards the robot not only for completing the task, but for reasoning along the right path.
Adaptive thinking steps — LaST-R1's most cyberpunk feature: it automatically adjusts its "thinking time" to task difficulty. Picking up an apple? One mental simulation may suffice. Untangling a knot of threads? It enters seconds-long deep logical deduction.
A 99.8% win rate — On hellishly difficult robot manipulation benchmarks like LIBERO, LaST-R1 achieved near-perfect results.

3. Wired Viewpoint: From "Automatic" to "Autonomous"

This is the logical awakening of robotics.

LaST-R1 means we have finally channeled the powerful reasoning ability exemplified by DeepSeek-R1 into robots with steel bodies. When a robot no longer merely "stops at a red light" but can "understand the traffic logic behind the red light and the physical interplay of braking distance," AGI truly gains feet.

In the future, when you walk into your kitchen and see your robot butler staring blankly at a pile of dirty dishes, don't interrupt it. It isn't slacking off—it's running a physical simulation iterating tens of thousands of times per second in latent space, planning the perfect chore path for you.

---

📑 Paper Details

Title: *LaST-R1: Reinforcing Action via Adaptive Physical Latent Reasoning for VLA Models*
Authors: Y. Zhang, J. Lee, S. Tan, et al.
Submitted: April 30, 2026 (arXiv update May 2, 2026)
arXiv ID: 2604.28192
Core contribution: First to introduce Latent Chain-of-Thought into VLA models, with the LAPO algorithm optimizing physical reasoning via reinforcement learning—marking a major leap from intuitive reaction to autonomous physical reasoning in embodied AI.