Research map / Vision-language and multimodal / Robot action and navigation

LaST-R1: Reinforcing Robotic Manipulation via Adaptive Physical Latent Reasoning

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Memory and subgoal-based manipulation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LaST-R1, an RL post-training framework (LAPO) that jointly optimizes latent chain-of-thought reasoning and action generation in a VLA, plus an adaptive latent CoT length mechanism, reaching 99.9% on LIBERO from one-shot SFT warm-up.

Problem

Latent-reasoning VLA models are trained only by imitation learning on static demonstrations, limiting adaptability and generalization; existing online RL for VLAs optimizes only action space and ignores the reasoning process.

Approach

LAPO treats latent reasoning embeddings as implicit decision variables and uses a step-level likelihood ratio (a Gaussian-based latent ratio plus action token ratio) so environment rewards shape both latent reasoning and actions; an adaptive <latent_end> token lets the policy choose its reasoning length.

Key results

• Achieves 99.9% average success rate on LIBERO with one-shot SFT warm-up, ranking first on all four suites (vs. pi_RL 98.3%).
• On real-world tasks, average success rises from 52.5% after warm-up to 93.75% after RL, versus 71.25% for full-SFT pi0.5.
• Ablation on LIBERO-Spatial: LAPO with latent reasoning reaches 99.9% vs. Action-Only + PPO 95.2%.

Surprising findings

• Adaptive latent CoT with the RL-optimized policy converges toward short reasoning lengths (2 or 4 tokens dominating), and the paper reports execution steps even below expert demonstrations in most scenarios.
• Setting the candidate-position count M=8 slightly degrades performance relative to M=4.

Limitations

• Real-world evaluation uses only four tasks with 20 rollouts per setting and human-judged success.
• Pre-training dataset is large but the LIBERO and real-world results depend on a one-shot or few-shot warm-up and specific hyperparameters (sigma, lambda weights).
• Latent-ratio derivation assumes isotropic Gaussian latents with fixed variance, an approximation.

Experimental setup

Datasets: LIBERO Open-X-Embodiment DROID RoboMIND

Benchmarks: LIBERO-Spatial LIBERO-Object LIBERO-Goal LIBERO-Long real-world Franka tasks

Metrics: success rate average rank out-of-distribution success rate

Baselines: OpenVLA GR00T-N1 pi0 pi0.5 OpenVLA-OFT GRAPE TGRPO VLA-RL SimpleVLA-RL RLinf-GRPO pi_RL

Builds on: PPO Qwen3-VL SimpleVLA-RL GRAPE DINOv3

Closest papers by meaning

• Robot-R1: Reinforcement Learning for Enhanced Embodied Reasoning in Robotics (NeurIPS 2025 · similarity 0.90)
• SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
• Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · similarity 0.88)
• Plan-Seq-Learn: Language Model Guided RL for Solving Long Horizon Robotics Tasks (ICLR 2024 · similarity 0.88)
• Multi-Stage Manipulation with Demonstration-Augmented Reward, Policy, and World Model Learning (ICML 2025 · similarity 0.88)
• SOLD: Slot Object-Centric Latent Dynamics Models for Relational Manipulation Learning from Pixels (ICML 2025 · similarity 0.88)