LaST-R1: Teaching Robots to Imagine Before They Act with Physical Latent Reasoning

Authors: Chenyang Gu, Jialin Gao, Ziyu Guo, Siyuan Qian, Yinxi Wang, Peng Jia, Chi-Wing Fu, Shanghang Zhang, Pheng-Ann Heng arXiv: 2604.28192 Institutions: The Chinese University of Hong Kong, Peking University, Simplexity Robotics Project page: https://siriyep.github.io/last-r1/

The Problem: When "See-Act" Isn't Enough

Vision-Language-Action (VLA) models such as RT-2, OpenVLA, and π₀ map visual observations and language instructions directly to robot actions. But they face a fundamental tension between reasoning (which requires time and imagination) and control (which requires speed).

Existing workarounds are compromised:

• Explicit linguistic reasoning (generating text like "move left, then grip the handle") is interpretable but slow, and discrete language crudely approximates continuous physical reality.
• Continuous latent reasoning is more expressive but has been trapped in static imitation learning — copying demonstrations without exploration or adaptation.
• Reinforcement learning (RL) offers exploration, but current RL methods for VLAs optimize only the action space, never the reasoning process itself.

LaST-R1's Three Innovations

1. Latent Chain-of-Thought Grounded in Physical Reality

LaST-R1 generates a sequence of latent reasoning tokens autoregressively, explicitly anchored on global future representations from a vision foundation model (VPT, Visual Proprioceptive Transformer). These tokens encode an imagined future — object trajectories, contact points, forces — and condition action generation. The result is a non-verbal "mental rehearsal" of physical dynamics in the model's latent space.

2. LAPO: Reinforcing Reasoning, Not Just Actions

Latent-to-Action Policy Optimization (LAPO) treats latent reasoning tokens as implicit decision variables. Reward flows back to update both the reasoning process and action generation, so failures can be traced to incorrect physical imagination rather than merely bad trajectories. LAPO computes a joint step-level likelihood ratio over latent and action tokens:

∇J = E[ Σ (advantage_t × ∇log π(latent_t, action_t | state_t)) ]

Joint optimization creates a feedback loop: better reasoning yields better actions, and successful actions reinforce the reasoning that produced them.

3. Adaptive Reasoning Length

An adaptive latent chain-of-thought mechanism lets the model emit a <latent_end> token when it has reasoned enough. Simple reactive tasks get brief reasoning; complex multi-step manipulation extends the chain. The optimal reasoning horizon is learned via RL, not manually programmed.

Results

• LIBERO benchmark: 99.8% average success rate across all four suites (Spatial, Object, Goal, Long). On LIBERO-Long, 99.4% with just one demonstration per task, versus ~97% for previous full-dataset state of the art.
• Convergence: LAPO reaches near-optimal performance in far fewer training steps than PPO baselines; latent reasoning acts as a "cognitive buffer" smoothing the RL optimization landscape.
• Real-world robots: Up to 44% improvement after LAPO post-training, ~90% average success on single-arm and dual-arm tasks (precise insertion, tool use, articulated objects), with zero-shot generalization to unseen objects, backgrounds, and lighting.

Why It Matters

LaST-R1 reframes robotic intelligence around *physical imagination* rather than a bolted-on cognition stage — closer to how humans reach for a cup by simulating the grasp before moving. Implications:

• Industrial robotics: one-shot warm-up plus RL refinement could cut per-task deployment from months to hours.
• Home robotics: physical reasoning plus trial-and-error adaptation suits unstructured environments.
• AI research: latent reasoning may extend beyond robotics, challenging the assumption that language is the only medium of thought.

Honest Critique

• Scale: the backbone is Qwen3-VL-4B; scaling to much larger models is unproven, and autoregressive latent generation plus parallel decoding raises real-time deployment concerns.
• Sim-to-real: real-world validation covers only a few controlled tasks; the gap persists for truly unstructured settings.
• Interpretability: latent reasoning is expressive but opaque — we can't easily ask the robot "what were you thinking?"
• Reward design: LAPO relies on clear success/failure signals; ambiguous or delayed rewards remain an open problem.

References

• Gu, C., et al. (2026). LaST-R1: Reinforcing Action via Adaptive Physical Latent Reasoning for VLA Models. arXiv:2604.28192.
• Project page: https://siriyep.github.io/last-r1/
• Liu, Z., et al. (2026). LaST₀: Latent Spatio-Temporal Chain-of-Thought for Robotic Vision-Language-Action Model. arXiv:2601.05248.
• Liu, J., et al. (2025). What Can RL Bring to VLA Generalization? An Empirical Study. arXiv:2505.19789.
• Schulman, J., et al. (2017). Proximal Policy Optimization Algorithms. arXiv:1707.06347.