Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.06271

VLA-ZO: Fast Zeroth-Order Adaptation for Vision-Language-Action Models

Jaemin Kim·Jiahn Kim·Taesik Gong
View original source ↗Read source document ↗Export metadata ↓

Abstract

Adapting vision-language-action (VLA) models to deployment-time distribution shifts is important for reliable robotic operation, but conventional first-order adaptation can exceed the memory budget of inference-oriented deployment platforms. Zeroth-order (ZO) optimization offers a forward-only alternative with inference-level memory, but accurate gradient estimation requires many perturbation queries, making naive ZO prohibitively slow for large VLA models. We present VLA-ZO, a framework for fast ZO adaptation that exploits the structure of VLA computation. By confining adaptation to the action side, VLA-ZO keeps the expensive vision-language prefix frozen and reuses its conditioning states across perturbation queries and optimizer steps, while schedule-aware prefetching hides state-transfer overhead. On LIBERO camera-viewpoint shifts, VLA-ZO reduces end-to-end adaptation time by 25.59$\times$ at $q=16$ and 32.54$\times$ at $q=64$ relative to baseline ZO, while improving average task success from 48.27% without adaptation to 58.17% and 63.58%, respectively. These results show that making ZO faster can make larger query budgets practical, providing a promising path toward resource-efficient VLA adaptation on deployment platforms.

Explore related subjects

cs.ROcs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Jaemin Kim, Jiahn Kim, Taesik Gong. 2026-10-05. VLA-ZO: Fast Zeroth-Order Adaptation for Vision-Language-Action Models. https://arxiv.org/abs/2610.06271

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

KEEP EXPLORING

Related papers

arXiv2026

A Three-Stage Offline SDRE-Based Control Framework for Human Motion Reproduction on a Suspended Bipedal Robot

This paper presents a three-stage offline command generation framework for reproducing human lower-limb motion on a suspended bipedal robot while matching torque trajectories computed from the robot dynamic model. First, State-Dependent Riccati Equation (SDRE) control derives the reference torque trajectory for the measured motion. Second, parameterized optimization converts this trajectory into trapezoidal joint velocity commands under motor speed and acceleration limits. Third, a proportional-integral-derivative linear quadratic regulator (PID-LQR) compensation scheme refines these commands using experimental tracking data. The platform executes the resulting profiles to reproduce human walking and squatting motions recorded by a Vicon system, allowing evaluation of tracking accuracy and repeatability. Results show that the average root mean square error (RMSE) and standard deviation (STD) of joint angles across repeated trials remain below 7° and 0.33°, respectively. Joint angle and torque trajectory comparisons show lower maximum RMSE and STD values than those for MPC and IPSO-PID in every reported case. The framework enables accurate and repeatable motion reproduction within actuator limits, providing controlled and measurable conditions that can reduce reliance on human participation and associated risks during preliminary evaluation of devices for assistive walking, gait training, and rehabilitation.

cs.RO↗
arXiv2026

LHM-Humanoid: Long-Horizon Human Motion Control for Continuous Object Transport in Cluttered Scenes

Physics-based human motion control can make a simulated character walk, sit, and manipulate objects with high physical realism. Almost always, though, this happens in short, isolated clips that are re-initialized between interactions. We instead aim for continuous, reset-free long-horizon motion: a physically simulated humanoid that repeatedly walks to a displaced object, lifts it with a balanced whole-body posture, carries it past obstacles, and places it at a goal, over and over within a single uninterrupted take. The hard part is not any individual motion but the transitions between them. Without a reset, each cycle must end in a state that both leaves the object just placed undisturbed and lets the next cycle begin, yet every placement leaves the character off-balance in a non-canonical pose where naive end-to-end reinforcement learning fails. Our key idea is to treat this handoff as a two-sided problem of recoverability: the character must disengage from the object it just placed so the prior success is preserved, and settle into a state from which a balanced continuation exists. Instead of engineering a transition by hand, we learn to shape where each cycle ends so that it lands in this recoverable region. We introduce LHM-Humanoid. One goal-conditioned controller completes a fetch--carry--place cycle and, through a learned release-and-retreat behavior, steers its terminal state into this region; a second controller then takes over from the resulting state distribution. Both are regularized by an adversarial motion prior and distilled into a single goal-conditioned policy that runs the whole sequence as one reset-free rollout. Across 350 cluttered layouts spanning four room types, LHM-Humanoid produces far more successful and stable long-horizon motion than end-to-end RL, hierarchical RL, and prior physics-based human-scene-interaction methods, on both seen and unseen scenes.

cs.RO↗
arXiv2026

Learning from Hallucinating Critical Points for Navigation in Dynamic Environments

Generating large and diverse obstacle datasets to learn motion planning in environments with dynamic obstacles is challenging due to the vast space of possible obstacle trajectories. Inspired by hallucination-based data synthesis approaches, we propose Learning from Hallucinating Critical Points (LfH-CP), a self-supervised framework for creating rich dynamic obstacle datasets based on existing optimal motion plans without requiring expensive expert demonstrations or trial-and-error exploration. LfH-CP factorizes hallucination into two stages: first identifying when and where obstacles must appear in order to result in a near-optimal motion plan, i.e., the critical points, and then procedurally generating diverse trajectories that pass through these points while avoiding collisions. This factorization avoids generative failures such as mode collapse and ensures coverage of diverse dynamic behaviors. We further introduce a diversity metric to quantify dataset richness and show that LfH-CP produces substantially more varied training data than existing baseline. Experiments in simulation demonstrate that planners trained on a LfH-CP generated dataset achieves higher success rates compared to a prior hallucination method.

cs.RO↗