RADAR: Benchmarking Vision-Language-Action Generalization via Real-World Dynamics, Spatial-Physical Intelligence, and Autonomous Evaluation

Yuhao Chen; Zhihao Zhan; Xiaoxin Lin; Zijian Song; Hao Liu; Qinhan Lyu; Yubo Zu; Xiao Chen; Zhiyuan Liu; Tao Pu; Tianshui Chen; Keze Wang; Liang Lin; Guangrun Wang

arXiv · 2602.10980

RADAR: Benchmarking Vision-Language-Action Generalization via Real-World Dynamics, Spatial-Physical Intelligence, and Autonomous Evaluation

Yuhao Chen·Zhihao Zhan·Xiaoxin Lin·Zijian Song·Hao Liu·Qinhan Lyu·Yubo Zu·Xiao Chen·Zhiyuan Liu·Tao Pu·Tianshui Chen·Keze Wang·Liang Lin·Guangrun Wang

View original source ↗Read source document ↗Export metadata ↓

Abstract

VLA models have achieved remarkable progress in embodied intelligence; however, their evaluation remains largely confined to simulations or highly constrained real-world settings. This mismatch creates a substantial reality gap, where strong benchmark performance often masks poor generalization in diverse physical environments. We identify three systemic shortcomings in current benchmarking practices that hinder fair and reliable model comparison. (1) Existing benchmarks fail to model real-world dynamics, overlooking critical factors such as dynamic object configurations, robot initial states, lighting changes, and sensor noise. (2) Current protocols neglect spatial--physical intelligence, reducing evaluation to rote manipulation tasks that do not probe geometric reasoning. (3) The field lacks scalable fully autonomous evaluation, instead relying on simplistic 2D metrics that miss 3D spatial structure or on human-in-the-loop systems that are costly, biased, and unscalable. To address these limitations, we introduce RADAR (Real-world Autonomous Dynamics And Reasoning), a benchmark designed to systematically evaluate VLA generalization under realistic conditions. RADAR integrates three core components: (1) a principled suite of physical dynamics; (2) dedicated tasks that explicitly test spatial reasoning and physical understanding; and (3) a fully autonomous evaluation pipeline based on 3D metrics, eliminating the need for human supervision. We apply RADAR to audit multiple state-of-the-art VLA models and uncover severe fragility beneath their apparent competence. Performance drops precipitously under modest physical dynamics, with the expectation of 3D IoU declining from 0.261 to 0.068 under sensor noise. Moreover, models exhibit limited spatial reasoning capability. These findings position RADAR as a necessary bench toward reliable and generalizable real-world evaluation of VLA models.

Explore related subjects

cs.RO

Keep this discovery

BibTeX RIS

Yuhao Chen, Zhihao Zhan, Xiaoxin Lin, Zijian Song, Hao Liu, Qinhan Lyu, Yubo Zu, Xiao Chen, Zhiyuan Liu, Tao Pu, Tianshui Chen, Keze Wang, Liang Lin, Guangrun Wang. 2026-02-11. RADAR: Benchmarking Vision-Language-Action Generalization via Real-World Dynamics, Spatial-Physical Intelligence, and Autonomous Evaluation. https://arxiv.org/abs/2602.10980

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

KEEP EXPLORING

Related papers

arXiv2026

IMLE-VLA: Fast Single-Step Action Generation for Vision-Language-Action Policies

Vision-language-action (VLA) policies leverage pretrained vision-language backbones to achieve strong cross-task generalization. A leading design couples this backbone with a dedicated continuous action head trained via diffusion or flow matching. However, such heads rely on iterative multi-step sampling, for example 10 Euler steps in $\pi_{0.5}$. This creates an inference bottleneck that produces stop-and-go movement in the robot and slower task completion. We introduce IMLE-VLA, which replaces the iterative action head with a single-step conditional generator trained via conditional Implicit Maximum Likelihood Estimation (cIMLE). The cIMLE objective promotes multimodal action coverage, avoiding the mode collapse of naive regression heads while eliminating multi-step sampling entirely. When IMLE-VLA is applied to $\pi_{0.5}$, it increases inference frequency 3.67x (55 Hz vs. 15 Hz), enabling up to 11x higher action throughput. On the 40-task LIBERO benchmark, IMLE-VLA achieves the highest average success rate (98.0%) among all baselines while leading in inference frequency. Under the test-time perturbations of LIBERO-plus, IMLE-VLA retains $\pi_{0.5}$'s robustness while other baselines degrade sharply, confirming that the cIMLE head preserves generalization. Real-world experiments on a Franka Emika Panda across four tasks demonstrate smoother motion (2.2x to 3.0x lower jerk) and faster task completion, with IMLE-VLA outperforming $\pi_{0.5}$ on every task and reducing average VLA inference time per episode by 3.9x to 6.6x. Videos and code are available at https://kianhk6.github.io/IMLE-VLA/

cs.RO↗

arXiv2026

ObstaDiff: Generalizable Diffusion Policy Learning via Obstacle-aware Representations

Imitation learning has achieved impressive results in robotic manipulation, yet most existing approaches assume clean backgrounds and lack explicit mechanisms for obstacle-aware motion generation. Extending such policies to cluttered, real-world scenes with unstructured obstacles remains a key generalization challenge. We present ObstaDiff, a decomposed diffusion-policy framework with a lightweight obstacle-aware visual encoder. ObstaDiff extracts a structured target-obstacle-background representation, enabling the downstream alignment policy to generate end-effector trajectories toward a target-centered bottleneck pose while reasoning about surrounding obstacles. We evaluate ObstaDiff on 61 real-robot greenhouse trials per method (366 executions in total). ObstaDiff achieves 75.41% average task success and 8.20% average obstacle collision rate, outperforming representative imitation-learning baselines and improving generalization in cluttered agricultural scenes.

cs.RO↗

arXiv2026

Testing Between the Test Cases: Proving End-to-End Steering in Conditions You Never Drove

AI-based automated vehicle testing is challenging because a model that passes every test condition can still fail in the real world. Formal verification offers a way to directly address this gap. On a simulated highway and an arterial road we trained two small end-to-end steering networks each in CARLA, one on clear conditions alone and one on clear, fog, night and low sun. All four models were driven against a 2.19 ft lane-departure budget. Without driving again, we used bound propagation, a formal method that reads the trained weights, to compute how far steering can drift at every disturbance strength between two captured images. One calculation covers more than a campaign could drive: on the arterial it spans 133 poses, where ten intensities each would be 10^133 combinations, in minutes on one GPU. Not only did formal verification find conditions that broke the clear-trained policy without simulation testing, it provided some preliminary evidence for potential failures between the test cases. Our overall conclusion is that formal verification is a viable complement to simulation, and could be adopted as a part of verification and validation for automated driving.

cs.RO↗