LaST-R1: Unified Vision–Language–Action Framework

• LaST-R1 is a unified vision–language–action framework that integrates adaptive latent Chain-of-Thought reasoning with reinforcement learning to improve complex robotic manipulation.
• It employs a two-stage latent reasoning process, where tokens generated in sequence are aligned with high-dimensional DINOv3 embeddings to predict short-term physical dynamics.
• The LAPO algorithm jointly optimizes latent and action sequences, achieving state-of-the-art performance (≈99.8% on LIBERO) and robust generalization to real-world challenges.

LaST-R1 is a unified Vision–Language–Action (VLA) framework that integrates continuous, adaptive latent Chain-of-Thought (CoT) reasoning over physical dynamics, coupled with a reinforcement learning (RL) paradigm designed to optimize both internal reasoning processes and action generation. This approach addresses limitations of static imitation learning and vanilla action-space RL, particularly for complex long-horizon robotic manipulation tasks. Its core components include a latent reasoning module, the Latent-to-Action Policy Optimization (LAPO) algorithm, and an adaptive mechanism for adjusting reasoning depth, resulting in state-of-the-art performance on the LIBERO benchmark and robust generalization to real-world and out-of-distribution (OOD) scenarios (Chen et al., 30 Apr 2026).

1. Vision–Language–Action Model Setting and Motivations

A VLA model π_θ receives multimodal input—visual observations and language instructions—and outputs a sequence of continuous controls in SE(3) for robotic effectors. For single-arm robots, each action at time step t is a 7-DoF real vector aₜ ∈ ℝ⁷, whereas dual-arm setups augment this to 14-DoF. Initial policy training employs supervised fine-tuning (one or few expert demonstrations), followed by RL-based policy refinement via interaction.

Prior work in this domain predominantly relied on static imitation learning, which is limited by the availability of high-quality demonstrations and lacks closed-loop correction, leading to compounding errors. Conversely, standard RL methods have optimized policies directly in action space, neglecting latent internal reasoning, resulting in brittle policies on difficult, long-horizon tasks. LaST-R1 is designed to bridge this gap by interweaving internal reasoning and control optimization.

2. Latent Chain-of-Thought Reasoning Module

LaST-R1 augments a vision–language transformer backbone (e.g., Qwen3-VL-4B with SigLIP2 encoder) with a two-stage decoding process. In the first stage, the model autoregressively generates a sequence of N_z latent reasoning tokens (z_{t,1}, …, z_{t,N_z}) that encode predictions about future physical states over a short horizon. Emission of a special <latent_end> token triggers the second stage, where the H-step action sequence is decoded in parallel.

Latent tokens are anchored to pre-computed DINOv3 <CLS> embeddings (4096 dimensions), projected into the model’s hidden size via top-k channel selection (resulting in a 2560-dimensional representation). Formally, for a multimodal state sₜ at timestep t, the latent generation at each position k follows:

$$z_{t,k} \sim \pi_\theta(z_{t,k} | s_t, z_{t,<k}), \quad k = 1, ..., N_z.$$

During supervised pretraining, the model aligns these outputs to offline DINOv3 targets $z_t^\star$ using a cosine similarity loss:

$$L_{\text{latent\_SFT}} = -\sum_{k=1}^{N_z} \cos(z_{t,k}, z_{t,k}^\star).$$

In RL post-training, behavioral changes in latents are regularized using an isotropic Gaussian likelihood, facilitating computation of a latent likelihood ratio for policy gradient estimation. Action generation is conditioned on the latent reasoning tokens, with deterministic or adaptive determination of the reasoning horizon.

3. Latent-to-Action Policy Optimization (LAPO)

LAPO is a novel RL algorithm developed for LaST-R1 that enables the joint optimization of both latent reasoning sequences (Zₜ) and action sequences (Cₜ) at each timestep. The RL objective is to maximize expected cumulative reward:

$$J_\text{RL}(\theta) = \mathbb{E}_{\tau \sim \pi_\theta} \left[ \sum_{t=0}^T \gamma^t r_t \right]$$

where τ denotes trajectory tuples $(s_0, Z_0, C_0, r_0), (s_1, Z_1, C_1, r_1), ...$. Policy updates use step-level likelihood ratios for both latent and action spaces:

$$\begin{align*} r_t^a(\theta) &= \exp(\log \pi_\theta(C_t | \cdot) - \log \pi_{\theta_\text{old}}(C_t| \cdot)) \ r_t^z(\theta) &= \exp\left(-\frac{1}{2\sigma^2} \sum_{k=1}^{N_z} \|z_{t,k}^{\text{old}} - z_{t,k}^\theta\|^2 \right) \end{align*}$$

Advantages $\hat{A}_t$ are computed via Generalized Advantage Estimation (GAE). The surrogate clipped policy loss combines likelihood ratios for both modalities:

$$L_{\text{policy}}(\theta) = -\mathbb{E}_t\left[\min_{m \in \{a,z\}} (r_t^m(\theta) \hat{A}_t, \text{clip}(r_t^m(\theta), 1-\epsilon, 1+\epsilon) \hat{A}_t)\right]$$

Combined with value estimation and auxiliary losses, this joint optimization allows external rewards to shape both cognitive (latent) reasoning and low-level control.

Empirically, the inclusion of reasoning tokens as an intermediary between perception and action smooths the optimization landscape, speeds convergence, and consistently improves success rates, especially for long-horizon tasks (achieving ≈99.8% on LIBERO suite; see below).

4. Adaptive Latent CoT Mechanism

To address the varying complexity of manipulation tasks, LaST-R1 incorporates an adaptive mechanism that enables dynamic adjustment of the latent reasoning horizon