Searcharxiv⌕ Search

arXiv subjects

Zsolt Kira

Publications and source records attributed to Zsolt Kira.

At least 19 recordsLinked to original sources

Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification

Verifiers--functions assigning rewards to agent behavior--have been key to AI progress in math, code, and games. However, extending gains to domains without clear-cut success criteria (e.g., computer use) remains a challenge: while humans can recognize desired outcomes, translating this intuition into scalable rules is nontrivial. Multimodal LLMs (MLLMs) offer a promising solution, given their world knowledge, human-preference alignment, and reasoning skills. We evaluate MLLM verifiers across web navigation, computer use, and robotics, spanning 13 models, 28 judge designs, diverse agent trajectories, and distinct downstream applications. We identify a critical limitation: a strong tendency for MLLMs to over-validate agent behavior, a phenomenon we term agreement bias. This bias is pervasive across models, is resilient to test-time scaling, and harms pipelines relying on MLLM judgments (e.g., self-improvement). We discuss principles for design and evaluation of verifiers, and introduce SGV, a method that improves verification by harnessing MLLMs' own capabilities via modulation of (un)conditional generation. First, an MLLM generates broad priors about desired behavior, independent of the data under evaluation. Then, conditioned on generated priors, it reasons over and judges a candidate trajectory. Our methods yield more human-aligned verifiers, with gains of up to 25pp in failure detection, 14pp in accuracy, and benefits extending to downstream applications. In self-improvement and online supervision, our verifiers boost task completion of a GUI specialist in OSWorld, a diffusion policy in robomimic, and a ReAct agent in VisualWebArena--surpassing the previous SOTA by 20pp. Finally, we release an updated version of VisualWebArena, featuring improved evaluators, high-fidelity environment parallelism with >10x speedups, and VWA-Lite, a 1/3-scale subset with comparable evaluation fidelity.

cs.AI↗

FlowEval: How Much Adaptation Does an Embodied Foundation Model Need at Deployment?

Large-scale pretraining is reducing the amount of adaptation required to deploy embodied foundation models, but it is not guaranteed that a pretrained model already provide the level of performance required in every deployment. Practitioners therefore need to understand how well a pretrained model matches a target deployment and how much adaptation data is left to be collected. We study whether a small set of successful target-domain demonstrations can reveal whether and how much target-domain data will be required before any adaptation. We introduce FlowEval, a set of training-free metrics that use the flow-matching objective to probe an embodied foundation model with a set of demonstrations. FlowEval quantifies predictive disagreement on successful target demonstrations to assess how well a pretrained model fits a target deployment. Across four embodied foundation model families: Pi-0.5, MolmoAct2, GR00T, and DreamZero, FlowEval distinguishes seen from unseen deployments across real-world and simulated manipulation. Further, by finetuning Pi-0.5 and GR00T on 17 RoboCasa365 manipulation tasks, we showed that FlowEval correlates with how success rate improves as additional adaptation data is provided. Together, these results suggest that FlowEval can serve as a lightweight, training-free diagnostic of how much model--deployment adaptation may be required.

cs.RO↗

AndroidFlux: Evaluating and Reward Modeling Failure Recovery Capabilities in Mobile Use Multimodal Agents

Mobile-use multimodal agents are typically evaluated by task success from a clean initial state, which does not isolate their ability to recover from intermediate errors. Because mobile environments change dynamically, a general mobile-use agent must complete tasks robustly even from unexpected states and in unseen environments. This requires knowing when and where an agent makes errors, and whether it can recover from errors and subsequent states. To evaluate this capability, we introduce AndroidFlux, a failure-recovery benchmark built on AndroidWorld. AndroidFlux restores recorded trajectories to controlled checkpoints before, at, and after the first recoverable error, so that different agents continue from the same device state and history. On these checkpoints, a complementary reward-modeling benchmark evaluates recovery-action preferences using downstream rollout outcomes. We further release an offline reinforcement learning (RL) dataset that improves both error recovery and overall task success. Our evaluations show that recovery performance only partially correlates with clean-start success, and that explicit reasoning helps only models trained with GUI-domain specific RL.

cs.LG↗

SeeTraceAct: Visibility-Aware Latent Planning from Cross-Embodiment Demonstration Videos

Vision-language-action models (VLAs) are promising general-purpose robot policies, but adapting them to new tasks typically requires costly task-specific teleoperation data. As an alternative, we study one-shot demo-conditioned VLAs, where a robot policy is conditioned on a single demonstration video of an unseen task. We find that existing end-to-end approaches often struggle when successful execution requires precisely localizing small target regions. To address this limitation, we propose SeeTraceAct, a demo-conditioned VLA framework that encourages precise spatial grounding through visibility-aware prediction of future end-effector traces. To enable reproducible evaluation with cross-embodiment demonstrations, we introduce and release RoboCasa-DC, a demo-conditioned extension of RoboCasa that pairs its robot trajectories with newly collected humanoid demonstrations. Experiments on RoboCasa-DC and a real-world benchmark, where a Franka Panda arm is conditioned on human demonstrations, show that SeeTraceAct outperforms baselines, achieving the best success rate across all four RoboCasa-DC settings and improving real-world average success by 9.5 percentage points.

cs.RO↗

Splat2Act: From Captured Scenes to Personalized VLA Manipulation

A user who brings a generalist manipulation policy into their home or office wants it to pick up their objects from their shelves, counters, and tables. Yet even a strong generalist like $π_{0.5}$ achieves 0% pickup success in our captured deployment scenes without adaptation, and collecting real demonstrations in every user's environment is prohibitively expensive. We present Splat2Act, an automated real-to-sim-to-real pipeline that personalizes a pretrained vision-language-action (VLA) policy to a specific scene and object set using only a handheld phone scan and images of the user's objects. Unlike prior real-to-sim work that reconstructs individual tabletops, Splat2Act operates on building-scale indoor environments: it builds a photorealistic digital twin with 3D Gaussian Splatting and collision geometry, automatically discovers tabletop and non-tabletop receptacles, constructs feasible pickup tasks, and generates successful demonstrations with a privileged simulation oracle, with only scene and object capture required from a human. Applied to 101 scanned campus environments with a median navigable area of 104.87 m$^2$ and 7 receptacles, the pipeline yields 852 receptacles and over 68,000 pickup demonstrations. We fine-tune $π_{0.5}$ in two stages: Stage 1 adapts to hundreds of public object assets, drawn either from the deployment scene alone or from all 101 scenes; Stage 2 personalizes to the user's own objects on the deployment scene's receptacles. Our strongest personalized policy reaches 62.5% aggregate pickup success in the target scene. Deployed on a real Stretch 3, the same policy achieves 63.9% success on a 12-task subset, closely tracking its 66.7% simulated success on those tasks. These results suggest that automated real-to-sim-to-real training can personalize VLA manipulation policies using only simulation-oracle data.

cs.RO↗

Beyond Interaction Capacity: Estimator Scaling with Recursive Models for CTR Prediction

Click-Through Rate prediction, a core task in recommendation and advertising systems, relies on modeling interactions among sparse categorical features. Explicit cross networks are a central paradigm for CTR prediction, and recent progress has largely come from increasing the interaction capacity of a single predictor through deeper cross networks and more expressive cross operators. We revisit whether continually increasing interaction capacity remains the most effective way to improve predictive performance, and find that its benefits quickly exhibit diminishing returns even as capacity continues to grow. This motivates a complementary scaling direction that we call estimator scaling, where additional resources are used to incorporate multiple related estimators rather than only enlarging a single predictor. Through theoretical analysis, we show that the gains from estimator scaling are governed by the amount of non-shared predictive variation available across estimators. However, exploiting this variation naively can be expensive: independently trained models provide substantial estimator diversity but require deployment cost to grow with ensemble size. This motivates a parameter-efficient realization of estimator scaling that can incorporate diversity from multiple estimator sources without maintaining multiple full models. Building on this view, we introduce RECursive Averaged Predictor (RECAP), a parameter-efficient recursive CTR model that operationalizes estimator scaling at three levels: distillation across independently trained models, exponential moving averaging over training trajectories, and aggregation over inference-time routes within a weight-shared recursive backbone. Experiments across multiple benchmarks establish new state-of-the-art predictive performance on standard benchmarks, while placing the RECAP on a favorable performance-parameter Pareto frontier.

cs.LG↗

SafeManip: A Property-Driven Benchmark for Temporal Safety Evaluation in Robotic Manipulation

Robotic manipulation is typically evaluated by task success, but reaching the correct final state does not guarantee safe execution. A robot may succeed despite contamination, premature release, or incorrect sequencing. These failures are poorly captured by task-completion metrics or isolated state checks because safety often depends on how behavior unfolds over time. We introduce SafeManip, a benchmark for evaluating temporal safety directly from low-level manipulation rollouts across eight physical and semantic categories. SafeManip grounds symbolic manipulation predicates into reusable temporal properties over finite executions, formalized with LTLf. We further introduce ManipVerse, a property ontology enabling semantic safety rules generalize across objects, tasks, and environments. Finally, SafeManip introduces per-trigger safety metrics that normalize violations by the safety-relevant opportunities that activate each obligation, accounting for differences in how often those opportunities arise across policies and tasks. We instantiate SafeManip on RoboCasa365 and LIBERO to audit seven state-of-the-art robot foundation-model policies. Our results reveal a gap between task capability and temporal safety: higher task success does not necessarily imply safer execution, successful rollouts can still contain violations, and failure patterns vary systematically across safety categories, task horizons, and manipulation suites. SafeManip provides a reusable layer for evaluating how safely manipulation tasks are executed, not only whether they are completed.

cs.RO↗

CAPEX: Efficiently Distilling Foundation Model Behavior into Deployable Robot Policies through Experience-Adaptive Reasoning

Robot learning has largely relied on human-teleoperated demonstrations to acquire effective learnable behaviors. However, human-operated data collection processes can be unintuitive, difficult to scale, and inherently asynchronous. We explore an alternative: distilling physical behavior from general-purpose multimodal foundation models into deployable robot policies by using the foundation model itself as an autonomous demonstrator. While sufficiently capable models can generate successful zero-shot manipulation trajectories, repeatedly invoking them during physical execution is slow and expensive, limiting their utility as scalable data generators. As a solution, we introduce CAPEX, an experience-conditioned demonstration collection framework that uses execution experience from previous attempts to adapt how frequently the foundation model must observe, reason, and replan. We evaluate across RoboCasa tasks and on physical Franka and bimanual YAM-arm platforms, measuring task success, model calls, token usage, collection time, and cost. We further train Diffusion Policy and ACT on matched sets of human-teleoperated and foundation-model-generated demonstrations to evaluate the downstream learning value of autonomously collected data. We find that CAPEX increases the number of successful demonstrations by 4.3x while reducing the cost per successful demonstration by 80%. Policies trained on CAPEX-generated data approach the performance of those trained on matched human demonstrations; with longer training, this gap largely closes for policies trained from scratch. These results suggest that foundation models can serve as scalable sources of reusable robot experience. Project page: https://capex-paper.github.io/

cs.RO↗

Hide-and-Seek in Trajectories: Discovering Failure Signals for VLA Runtime Monitoring

Vision-Language-Action (VLA) models enable robots to follow natural language instructions and generalize across diverse tasks, but they remain vulnerable to execution failures that compromise reliability in real-world deployment. Detecting such failures during execution is therefore critical for the robust deployment of embodied systems. Existing failure detection methods either rely on expensive action resampling or external models, while alternatives propagate trajectory-level labels uniformly across every timestep, obscuring localized failure signals. In this paper, we propose \textbf{Hide-and-Seek}, a framework that formulates VLA failure detection as a coarsely supervised learning problem. By combining inter-trajectory and intra-trajectory contrastive objectives, Hide-and-Seek localizes failure-indicative actions and induces temporally structured failure signals from trajectory-level supervision alone, without any step-level annotation. We evaluate Hide-and-Seek on LIBERO, VLABench, and a real-world robotic platform across three representative VLA policies: OpenVLA, $π_0$, and $π_{0.5}$.Our method achieves state-of-the-art multi-task failure detection performance with a practical accuracy--timeliness trade-off under conformal prediction, and generalizes well to both seen and unseen tasks.

cs.RO↗

Opt2VLA: Force-Aware Vision-Language-Action for Contact-Rich Humanoid Whole-Body Manipulation

Humanoid robots are expected to perform diverse human-level tasks in daily environments, many of which require precise regulation of interaction forces. While recent vision-language-action (VLA) models have shown promise for semantic planning and visuomotor control, existing humanoid systems primarily represent actions through geometric motion goals and rely on whole-body controllers focused on motion tracking, with limited explicit reasoning or control of interaction forces. This limitation is particularly relevant in contact-rich tasks, where geometrically similar motions may require different force regimes depending on the task context and where visual observations may become unreliable after contact. In this work, we present Opt2VLA, a force-aware VLA framework that introduces explicit force commands at the VLA-to-control interface for humanoid whole-body manipulation. A single multi-task VLA policy jointly predicts both geometric motion goals and continuous contact-force references, which are tracked by task-specific reinforcement learning (RL)-based whole-body controllers. To provide scalable and physically grounded supervision, we generate dynamically feasible and contact-consistent training data via whole-body trajectory optimization (TO) with explicit force references. We evaluate Opt2VLA on three contact-rich humanoid tasks and show that explicit force conditioning enables more accurate and consistent force regulation than motion-only control, while physically grounded torque supervision from TO further improves force tracking accuracy and stability. Closed-loop evaluations further demonstrate language-conditioned force modulation with Opt2VLA in simulation and on humanoid hardware.

cs.RO↗

Lies We Can See: Joint Verbal and Non-Verbal Deception by VLM Agents in Embodied Social Interactions

Strategic deception by LLM and VLM agents has emerged as a central AI alignment and safety concern. Social-deduction games (where each player holds a hidden role and communicates with others to deduce identities) serve as the canonical testbed, particularly in multi-agent settings. Existing testbeds, however, are text-only and run on a single fixed agent configuration, missing the non-verbal sensorimotor channels treated as core by deception taxonomies and leaving it ambiguous whether an observed behavior reflects the underlying model or the surrounding harness. We introduce MineAmongUs, a 3D multimodal Among Us sandbox where imposter agents must deceive crewmates through joint verbal and non-verbal action. We also propose ARIA, a configurable VLM-agent harness that exposes five cognitive-component ablation axes; and an atom- and arc-level annotation scheme grounded in deception taxonomies and operationalized at scale by an LLM-as-a-Judge reaching near-human atom-labeling agreement. Empirical results show that VLM agents pursue imposter wins through joint verbal and non-verbal deception, with non-verbal channels emerging as the more decisive winning contributors across both harness ablation and cross-VLM evaluation. Taken together, our work opens a new path for embodied VLM-agent alignment research.

cs.CL↗

K9-Bench: Evaluating Multimodal LLMs on Canine-Centric Videos

MLLMs have shown strong zero-shot capabilities across diverse inputs such as across images, video, audio, and text. A crucial, yet underexplored, application of these models lies in understanding and modeling animal-centric scenarios. As animals are integral to millions of households, benchmarking next-generation AI models on pet-focused tasks, ranging from recognizing distress signals to enabling responsive robotic companions, is essential for building AI systems that can work alongside us. We introduce K9-Bench, a novel benchmark focused on real-world domestic dog videos, specifically targeting canine action and interaction understanding via approximately 5000 question-answer pairs across 907 videos spanning 5 distinct task categories that test long-form, canine-centric multimodal reasoning in MLLMs. To create this dataset, we propose a scalable, VLM/LLM-powered data generation pipeline that automatically mines canine-centric videos from the web and curates QA pairs requiring fine-grained, multi-hop reasoning over canine actions and temporally extended interaction sequences. We implement bias mitigation strategies designed to eliminate biases introduced by VLMs during dataset curation. Through extensive experimentation, we find that frontier MLLMs exhibit limited zero-shot performance on canine-centric tasks: although state-of-the-art closed-source models outperform open-source counterparts, they still struggle with compositional reasoning over subtle posture and interaction cues spread over long horizons. We observe that generic chain-of-thought prompting provides only modest performance for such long-horizon reasoning. Beyond a novel dataset for canine activity analysis, K9-Bench provides a general-purpose dataset construction pipeline that can be adapted to other low-data domains for quantitative analysis. Our project website is available at: https://ogmenrobotics.github.io/K9Bench.

cs.CV↗

Video2Sim2Real: Full-Stack Autonomous Dexterous Skill Acquisition from a Single Human Video

Human manipulation videos are a convenient and intuitive source for robot learning. However, directly transferring human dexterity to robots remains challenging due to perception errors and embodiment gap. To address this, we introduce Video2Sim2Real, a full-stack framework for autonomous skill acquisition from a single human manipulation video. Our framework first uses off-the-shelf foundation models to reconstruct a simulator-ready digital twin and extract robot and object motion priors. Rather than treating the extracted robot motion as a reliable reference throughout execution, our key idea is to recover and leverage the most fundamental sources of supervision from the demonstrated skill: We identify object-centric keyframes to optimize the corresponding robot configurations using object information from the simulator, and use these configurations as anchors that refine the robot motion such that it ultimately has the desired impact on the environment. To bridge the remaining sim-to-real gap, we introduce a sim-to-real strategy that decouples robustness to noisy and incomplete perception from variations in hand-object interaction dynamics. Specifically, we learn to recalibrate robot configurations from noisy real-world point clouds via IL, and leverage residual RL to perform local finger-level adaptations to ensure for robust and effective interactions. Finally, a collision-aware motion planning module enables spatial generalization to novel object configurations. Across several everyday manipulation tasks, Video2Sim2Real improves simulated task success, safety, and trajectory coherence over numerous baselines, and achieves better sim-to-real transfer than existing techniques. These results demonstrate a promising path toward autonomous dexterous skill acquisition from human videos.

cs.RO↗

EVE: A Generator-Verifier System for Generative Policies

Visuomotor policies based on generative such as diffusion and flow-matching have shown strong performance for robotics applications but degrade under distribution shifts, demonstrating limited recovery capabilities without costly finetuning. In the language modeling domain, test-time compute scaling has revolutionized the reasoning capabilities of modern LLMs by enabling candidate solution refinement. These methods typically leverage foundation models as verification modules in a zero-shot manner to score candidate solutions. We hypothesize that generative policies can similarly benefit from additional inference-time compute that employs zero-shot VLM-based verifiers in a generation-verification framework. To this end, we introduce EVE: a modular, generator-verifier interaction framework that boosts the performance of pretrained generative policies at test time, with no additional training. EVE wraps a frozen base policy with multiple zero-shot, VLM-based verifier agents. Each verifier proposes action refinements to the base policy candidate actions, while an action incorporator uses classifier guidance to fuse aggregated verifier feedback into action denoising. We study design choices for generator-verifier information interfacing across a system of verifiers with distinct capabilities. Across diverse simulated and real robotic tasks and embodiments, EVE consistently improves success rates without additional policy or verifier training. Through extensive ablations, we isolate the contribution of verifier capabilities and action incorporator strategies, offering practical guidelines to build scalable, modular generator-verifier systems for embodied control.

cs.RO↗

The Geometry of Robustness: Optimizing Loss Landscape Curvature and Feature Manifold Alignment for Robust Finetuning of Vision-Language Models

Fine-tuning approaches for Vision-Language Models (VLMs) face a critical three-way trade-off between In-Distribution (ID) accuracy, Out-of-Distribution (OOD) generalization, and adversarial robustness. Existing robust fine-tuning strategies resolve at most two axes of this trade-off. Generalization-preserving methods retain ID/OOD performance but leave models vulnerable to adversarial attacks, while adversarial training improves robustness to targeted attacks but degrades ID/OOD accuracy. Our key insight is that the robustness trade-off stems from two geometric failures: sharp, anisotropic minima in parameter space and unstable feature representations that deform under perturbation. To address this, we propose GRACE (Gram-aligned Robustness via Adaptive Curvature Estimation), a unified fine-tuning framework that jointly regularizes the parameter-space curvature and feature-space invariance for VLMs. Grounded in Robust PAC-Bayes theory, GRACE employs adaptive weight perturbations scaled by local curvature to promote flatter minima, combined with a feature alignment loss that maintains representation consistency across clean, adversarial, and OOD inputs. On ImageNet fine-tuning of CLIP models, GRACE simultaneously improves ID accuracy by 10.8%, and adversarial accuracy by 13.5% while maintaining 57.0% OOD accuracy (vs. 57.4% zero-shot baseline). Geometric analysis confirms that GRACE converges to flatter minima without feature distortion across distribution shifts, providing a principled step toward generalized robustness in foundation VLMs.

cs.CV↗

SAGE: Sink-Aware Grounded Decoding for Multimodal Hallucination Mitigation

Large vision-language models (VLMs) frequently suffer from hallucinations, generating content that is inconsistent with visual inputs. Existing methods typically address this problem through post-hoc filtering, additional training objectives, or external verification, but they do not intervene during the decoding process when hallucinations arise. In this work, we introduce SAGE, a Sink-Aware Grounded Decoding framework that mitigates hallucinations by dynamically modulating self-attention during generation. Hallucinations are strongly correlated with attention sink tokens - punctuation or function tokens that accumulate disproportionate attention despite carrying limited semantic content. SAGE leverages these tokens as anchors to monitor grounding reliability in real time. At each sink trigger, the method extracts semantic concepts from the generated sequence, estimates their visual grounding using both self-attention maps and gradient-based attribution, and measures their spatial agreement. Based on this signal, self-attention distributions are adaptively sharpened or broadened to reinforce grounded regions or suppress unreliable ones. Extensive experiments across diverse hallucination benchmarks demonstrate that SAGE consistently outperforms existing decoding strategies, achieving substantial reductions in hallucination while preserving descriptive coverage, without requiring model retraining or architectural modifications. Our method achieves an average relative improvement of 10.65% on MSCOCO and 7.19% on AMBER across diverse VLM architectures, demonstrating consistent gains in hallucination mitigation.

cs.CV↗

Sim2real Image Translation Enables Viewpoint-Robust Policies from Fixed-Camera Datasets

Vision-based policies for robot manipulation have achieved significant recent success, but are still brittle to distribution shifts such as camera viewpoint variations. Robot demonstration data is scarce and often lacks appropriate variation in camera viewpoints. Simulation offers a way to collect robot demonstrations at scale with comprehensive coverage of different viewpoints, but presents a visual sim2real challenge. To bridge this gap, we propose MANGO -- an unpaired image translation method with a novel segmentation-conditioned InfoNCE loss, a highly-regularized discriminator design, and a modified PatchNCE loss. We find that these elements are crucial for maintaining viewpoint consistency during sim2real translation. When training MANGO, we only require a small amount of fixed-camera data from the real world, but show that our method can generate diverse unseen viewpoints by translating simulated observations. In this setting, MANGO outperforms all other image translation methods we tested. In certain real-world tabletop manipulation tasks, MANGO augmentation increases shifted-view success rates by over 40 percentage points compared to policies trained without augmentation.

cs.CV↗

Memo: Training Memory-Efficient Embodied Agents with Reinforcement Learning

To enable embodied agents to operate effectively over extended timeframes, it is crucial to develop models that form and access memories to stay contextualized in their environment. In the current paradigm of training transformer-based policies for embodied sequential decision-making tasks, visual inputs often overwhelm the context limits of transformers, while humans can maintain and utilize a lifetime of experience compressed as memories. Significant compression is possible in principle, as much of the input is irrelevant and can be abstracted. However, existing approaches predominantly focus on either recurrent models with fixed-size memory or transformers with full-context reliance. In this work, we propose Memo, a transformer-based architecture and training recipe for reinforcement learning (RL) on memory-intensive, long-horizon tasks. Memo incorporates the creation and retrieval of memory by interleaving periodic summarization tokens with the inputs of a model during training. We demonstrate Memo's effectiveness on a gridworld meta-RL benchmark and a multi-object navigation task in photo-realistic indoor settings. Memo outperforms naive long-context transformer baselines while being more compute and storage efficient. Additionally, Memo generalizes better to longer contexts at inference time and remains robust in streaming settings, where historical context must be truncated to fit inference constraints. Our code is available at: https://github.com/gunshi/memo.

cs.AI↗