Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.28166

Reasoning as a Distribution: Generative Plan Latents for Vision-Language-Action Models

Matteo Cercola·Katie Z Luo·Milan Ganai·John Alora·Simone Formentin·Marco Pavone
View original source ↗Read source document ↗Export metadata ↓

Abstract

Vision-Language-Action (VLA) policies increasingly reason before acting, but existing reasoning models compress this intermediate decision into a single representation. When several plans are valid, this commitment hides alternatives that could otherwise be selected or revised. We introduce the Generative Plan Latent (GPL), which represents the reasoning state as a low-dimensional conditional distribution over plans, such as which object, grasp, or route to use. This makes the reasoning state addressable: without retraining the VLA, plans can be sampled (coverage), selected on request (steering), and reweighted by deployment outcomes (outcome adaptation). We test all three on SimplerEnv manipulation with $π_0$, and extend the analysis to RoboCasa, LIBERO, and a maze domain. Steering the plan raises the fraction of executions that follow the request from 0.48 to 0.94, while the same guidance on the native action distribution shows no reliable gain. Finally, we show that an explicit plan is most useful when it carries decision information not already determined by the observation.

Explore related subjects

cs.RO

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Matteo Cercola, Katie Z Luo, Milan Ganai, John Alora, Simone Formentin, Marco Pavone. 2026-10-20. Reasoning as a Distribution: Generative Plan Latents for Vision-Language-Action Models. https://arxiv.org/abs/2610.28166

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Synergistic Action and Style Priors for Learning Legged Locomotion

Deep Reinforcement Learning (DRL) enables robots to learn complex behaviors through interaction with the environment. However, due to the unrestricted nature of the learning algorithms, the resulting solutions are often brittle and appear unnatural. This is especially true for learning direct joint-level torque control, as inductive biases are difficult to integrate into the learning process. We propose an inductive bias for learning locomotion that is especially useful for torque control: latent actions learned from a small dataset of expert demonstrations. This prior allows the policy to directly leverage knowledge contained in the expert's actions and facilitates more efficient exploration. We observe that the agent is not restricted to the reward levels of the demonstration, and performance in transfer tasks is improved significantly. Latent action priors combined with style rewards for imitation lead to a closer replication of the expert's behavior. Videos are available at https://sites.google.com/view/latent-action-priors.

cs.RO↗
arXiv2026

Kamino: GPU-based Simulation of Rigid Multi-Body Physics with Kinematic Loops

We present Kamino, a GPU-based physics solver for massively parallel simulations of heterogeneous highly-coupled mechanical systems. Implemented in Python using NVIDIA Warp and integrated into the Newton framework, it enables the application of data-driven methods, such as large-scale reinforcement learning, to complex robotic systems that exhibit strongly coupled kinematic and dynamic constraints such as kinematic loops. The latter are often circumvented by approximating the system's topology as a kinematic tree, and incorporating explicit loop-closure constraints or so-called mimic joints. Kamino aims at alleviating this burden by natively supporting these types of couplings. Moreover, Kamino supports heterogeneous worlds, allowing for batched simulation of structurally diverse robots on a single GPU. At its core lies a state-of-the-art constrained optimization algorithm based on ADMM, which computes constraint forces by solving the constrained rigid multi-body forward dynamics transcribed as a nonlinear complementarity problem. This leads to high-fidelity simulations that resolve contact dynamics without resorting to approximate models that simplify or convexify the problem. We demonstrate RL policy training on a biped with six nested kinematic loops, generating a feasible walking policy while simulating $4{,}096$ parallel environments on a single GPU.

cs.RO↗
arXiv2026

Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models

A fine-tuned Vision-Language-Action (VLA) policy can easily place the alphabet soup in the basket on command, just as it saw during fine-tuning. However, if the basket is shifted five centimeters to the left, it drops the soup off the table. We developed an interpretability pipeline to investigate what such a policy is computing. Injecting task A's activations into the action expert while the policy looks at task B's scene makes pi0.5 execute task A's motor trajectory in 99.6\% of episodes ($n{=}1{,}968$). The injected program is bound to absolute workspace coordinates and ignores the visible scene, which provides a mechanistic account of the perturbation brittleness that concurrent benchmarks report. The binding holds on a physical UR10e: directed replay redirects the commanded task in 26/26 rollouts in one direction and 18/26 in the other, with no-injection and random-direction controls at the floor, and once the source object is moved 10--15\,cm the arm grasps and releases at the recorded coordinates in 50/50 trials, picking up the wrong block when one sits there and closing on empty air when none does. Applying the same interventions to six VLAs on LIBERO, MetaWorld, SimplerEnv, and ALOHA over 420{,}000+ rollouts yields three further findings: every architecture encodes the instruction, and four of the five weight it below vision whenever the scene alone identifies the goal; under the standard replacement write-back the SAE pooling that preserves behavior is architecture-specific (pi0.5 per-token, SmolVLA indifferent), and goal selection and motor generation separate into VLM and expert pathways wherever the two are distinguishable. We release \textbf{Action Atlas} (https://action-atlas.com), an interactive platform covering all 520+ trained SAEs, 79 identified concepts, and the scored UR10e rollouts. Code: https://github.com/CWRU-AISM/action-atlas

cs.RO↗