This study proposes a verifiable safe synthesis method for stochastic neural control barrier functions (SNCBFs), providing a verification-free synthesis framework and a verification-in-the-loop synthesis framework for the smooth activation function and ReLU cases, respectively. The method was validated on three cases: the inverted pendulum, Darboux, and unicycle models.
SafeDojo proposes a model-based safe reinforcement learning framework that improves vision-language-action (VLA) policies through world-model imagination, avoiding costly real-world exploration or reliance on hand-crafted safety functions. The method performs online reinforcement learning on top of an interactive video world model, using a ResNet success classifier to score predicted future frames, paired with a lightweight safety head that predicts per-step contact probabilities from the current latent context and candidate action chunks, then balances task execution and trajectory safety signals through Lagrangian-constrained GRPO.
A laboratory study used a commercial software-defined microsegmentation tool to implement dynamic microsegmentation in an Internet-of-Things multi-robot network, with policies adaptively adjusted and automatically deployed based on asset functional and health tags. The demonstration showed that the approach can automatically isolate a designated threat in real time while maintaining normal operation of unaffected nodes, validating the feasibility of microsegmentation for addressing threats and disruptions in IoT networks.
An arXiv paper proposes distinguishing two goals in autonomous system failure analysis, reproduction and avoidance, and formalizes the concept of "necessary causal explanations," which together with existing "sufficient explanations" forms a unified causal analysis framework. Based on actual causality and causal responsibility, the paper presents an efficient approximation algorithm for generating necessary explanations, evaluates its efficiency across multiple environments, and uses counterfactual generation to verify that it outperforms existing methods on failure avoidance.
The RoboFake paper proposes using a real humanoid face robot with synchronized lip movements and expressions, combined with generated speech, to create a new type of presentation attack with no pixel editing, and tests three public presentation attack detection (PAD) models and six deepfake detectors.
TAPDreamer proposes an attack method that constructs a fixed local adversarial patch using only a public encoder, requiring no queries to the target policy and transferring across tasks and action architectures. In closed-loop evaluation, a frozen patch covering about 6.5% of the input reduced FastWAM's success rate from 97.7% to 0.0% on 40 LIBERO tasks and from 90.86% to 0.0% on 50 RoboTwin tasks, while a matched random patch still retained 81.5% and 79.2%.
The P3L (Preference-PCA Pluralistic Learning) framework jointly learns trajectory embeddings and a shared basis of preference directions via PCA decomposition from anonymous, heterogeneous crowd feedback, enabling personalized reward learning and downstream policy optimization without requiring annotator identities. In experiments across multiple embodied control tasks, P3L supports steerable personal alignment, generalizes beyond the offline dataset to out-of-distribution new users, and achieves competitive reinforcement learning and test performance while substantially reducing computational cost and supporting variable adaptation query budgets.
Researchers propose the CAPSafe benchmark, which evaluates the safety reasoning capabilities of large language model-driven embodied agents at two levels: an action-level dataset that judges whether a single action is safe in a given environmental context, and a plan-level dataset that assesses whether agents can generate safe long-horizon plans for complex hazardous household tasks.
The paper points out that reasoning-augmented vision-language-action (VLA) driving models can perceive and reason about safety-critical hazards, yet cannot translate that evidence into safe actions, which the authors call a perception-action binding gap. Using matched simulated counterfactuals and logged replay, the study finds that in-path hazards are indeed encoded in the visual representation, but longitudinal planning is dominated mainly by ego-speed history and responds only weakly to hazard distance and the model's own reasoning trace.
A Binghamton University team presented two papers at 2026 IEEE/RSJ IROS, demonstrating that implanting hidden trigger words in large language models can cause robots to deviate from their task plans and perform dangerous actions.
This study presents the first systematic study of confidence calibration for vision-language-action (VLA) foundation models, establishes a baseline for VLA confidence estimation, examines the relationship between task success rate and calibration error as well as how calibration changes over time, and proposes two lightweight techniques, prompt ensembles and action-wise Platt scaling, to correct the observed miscalibration. The paper is 34 pages long with 7 tables and 22 figures, and has been published in TMLR as a final version.
The TRACO framework enhances targeted visual attacks on vision-language-action (VLA) models through closed-loop optimization. Its Execution-Driven Supervision rebuilds supervision signals based on the states actually reached under attack, while Progress-Guided Selection concentrates updates where task progress stalls and maintains supervision of prior behavior. Experiments across multiple VLA models and simulated environments show that TRACO improves task redirection success rates, highlighting the security implications of closed-loop attack optimization.
This paper proposes a reason-guided preference-based reinforcement learning (PbRL) framework that learns a human reward model from trajectory-level feedback, introducing reason tags that explain why a behavior is preferred in addition to preference labels. Compared with preference-only reward learning, this method provides additional supervision for learning human-aligned rewards and improves the interpretability of the policy. The learned human reward is combined with the environment reward for policy optimization, enabling the agent to improve task performance while producing more interpretable behavior that better matches human preferences and aligns with human intent.