AI Bites | YouTube Channel

Can LLMs discover physical laws they were never asked to look for? LongChang explores this question by placing language models in simulated worlds where familiar rules—such as how objects fall—have been altered, then letting them freely investigate by manipulating objects, designing tools, and forming their own questions. The framework evaluates whether models can notice anomalies, design experiments, and support claims with evidence, all without human-provided research tasks or hints. Across 12 LLMs, GPT-6-Astra achieved a 16.7% discovery rate, while the other 11 did not discover the altered laws. The findings highlight both the promise and current limits of autonomous scientific exploration. Title: Can LLMs Uncover Hidden Physical Laws through Free Exploration? Project URL: zhenyangcs.github.io/LongCha… Paper: arxiv.org/abs/2610.37649v1

Following the same object across different camera views—and keeping its identity consistent over time—is challenging when appearance changes or occlusions cause frame-by-frame methods to drift. This work introduces XTrack, a framework that adapts a pretrained video diffusion transformer to transport information from an RGB video to object masks while jointly modeling cross-view matching and temporal tracking. It combines focal cross-view attention alignment, hard-negative conditioning, local context crops, and curated pseudo-labels from sparse Ego-Exo4D annotations to improve reliable grounding and dense video-level supervision. Title: Repurposing Video Diffusion Transformers for Cross-View Temporal Object Correspondence Project URL: cvlab-kaist.github.io/XTrack… Paper: arxiv.org/abs/2610.37633v1

Collecting high-quality robot manipulation data shouldn’t require expensive infrastructure or closed tracking systems. OpenJaws is an open hardware and software framework for sensorized Universal Manipulation Interfaces (UMIs), combining a modular 3D-printed gripper with a low-cost stereo camera to recover metric camera trajectories and improve end-effector localization. The paper introduces UMI-Loc-Bench-v1, a benchmark for evaluating localization accuracy in handheld data-collection devices, and shows that stereo localization outperforms monocular alternatives. Policies trained with OpenJaws data are also deployed on a real robot across multiple tasks. Hardware, software, and data are released under a commercially permissive license: open-jaws.github.io/ Title: OpenJaws: An Open Hardware UMI Framework with Stereo Localization for Generalist Manipulation Project URL: open-jaws.github.io Paper: arxiv.org/abs/2610.36598v1

What if you could stylize selected regions of a video—without changing everything else? LociArt introduces a unified framework for free-style local video stylization, explicitly controlling both where a style appears and which style is applied while preserving non-target content and maintaining temporal consistency. Its GRNStyle model combines style-aware masks with spatial geometry and spatiotemporal context, enabling precise regional editing and joint stylization of multiple regions with the same or different styles. The work also presents the LociArt-100K dataset and benchmark for evaluating regional style quality, non-target preservation, and temporal consistency. Title: LociArt: A Unified Paradigm for Free-Style Local Video Stylization Project URL: higherhu.github.io/lociart/ Paper: arxiv.org/abs/2610.36456v1

What does it take to build a profitable AI SaaS as a solo founder? Short answer: fully relate to the problem. Long answer: watch the full podcast here: youtu.be/_WkFDvKv7Fc It was lovely talking with the solo founder of @myreader_ai, @prasann_pandya, about his journey building, launching, and living the "indie hacker" life with his SaaS product. Hope it helps the indie hackers here!

What does it take to build a profitable AI SaaS as a solo founder? Short answer: fully relate to the problem. Long answer: watch the full podcast here: youtu.be/_WkFDvKv7Fc Was lovely talking with the solo founder of @myreader_ai @prasann_pandya about his journey building, launching, and living the "indie hacker" life with his SaaS product. #IndieHackers #founders #AI #SaaS

Robust reinforcement learning is essential for deploying reliable policies in the real world—but comparing methods and reusing implementations remains difficult. RobustRLlib addresses both challenges with a unified library and benchmark covering 17 representative online, offline, and safe RL algorithms. Its evaluation platform spans diverse real-world shifts, humanoid tasks, and vision-language-action settings, with systematic testing across 273 perturbation configurations. The resulting capability map highlights where current methods remain effective—and where important robustness gaps persist. Open source and designed for extensibility, RobustRLlib supports reproducible research, practical reuse, and the development of new robust RL algorithms. Title: RobustRLlib: A Unified Library and Benchmark for Robust Reinforcement Learning Algorithms Project URL: robustrl-lib.github.io Paper: arxiv.org/abs/2610.32971v1

Humanoid robots can learn diverse whole-body behaviors from ordinary monocular human videos—but turning those videos into motions a robot can actually execute remains a major challenge. ViT-H introduces a generation-and-tracking approach: a diffusion model directly generates humanoid reference motions from video, bypassing separately optimized human-motion reconstruction and retargeting stages. The system then uses motion-tracking feedback to refine the generator through direct preference optimization, improving the executability of generated motions. An autoregressive version supports real-time video teleoperation with sub-400 ms latency. Experiments on the Unitree G1 show improvements in motion quality and executability over reconstruction–retargeting–tracking pipelines, with performance scaling across model size, training data, and different motion trackers. Title: ViT-H: Generative Video-To-Humanoid Control with Motion Tracking Feedback Project URL: awfuact.github.io/vit-h/ Paper: arxiv.org/abs/2610.32966v1

Reconstructing a 3D scene from only a few widely spaced views is challenging because the unseen regions are uncertain. PixGR addresses this by combining feed-forward 3D Gaussian reconstruction with lightweight pixel-space generative modeling. The method encodes input views once into a shared scene representation, then uses a compact decoder to generate target views from noise while directly optimizing the underlying 3D Gaussian representation through differentiable rendering. By avoiding VAE-based latent compression, PixGR enables direct pixel-space supervision while preserving fast inference. Experiments show improved reconstruction in sparse-view and view-extrapolation settings, with sub-second inference and performance competitive with much heavier pretrained generative models. Title: PixGR: Pixel-Space Generative Decoding for Fast Feed-Forward 3D Reconstruction Project URL: ranrhuang.github.io/pixgr/ Paper: arxiv.org/abs/2610.31908v1

Compressing a large model is not enough—if its sparse structure makes inference slower, deployment gains can disappear. S-SALAAD addresses this challenge by jointly learning model weights, low-rank components, and hardware-compatible structured sparsity during training. A single checkpoint can then support multiple memory budgets, while dedicated CPU and GPU kernels exploit the learned structure to reduce memory use and inference latency. Evaluations on Llama 2 and Qwen3, across pretraining and distillation settings, show improved trade-offs among perplexity, memory, and latency relative to the evaluated baselines—supporting more practical large-model deployment on edge devices. Title: Learning Structured Sparsity for Large Models: Bridging Model Compressibility and Efficient Inference Project URL: egretwalker.github.io/S-SALA… Paper: arxiv.org/abs/2610.31897v1

How can robots learn from real-world interaction without turning mistakes into unsafe events? SEVI—Safe Evidential Real-World Reinforcement Learning—adds a reliability-aware safety layer to online policy learning. Before each action, SEVI estimates both the risk of future constraint violations and how much evidence supports that estimate. If an action appears unsafe or the prediction is too uncertain, an offline-trained recovery policy redirects the robot toward a lower-risk region—all in a single, low-latency forward pass. Evaluated on simulated manipulation tasks and a real-world USB insertion task, including distribution shifts, SEVI achieved higher final task success and fewer unsafe events than the evaluated baselines. It reduced unsafe events by 13.4–57.1% relative to recovery-based safe RL. Title: SEVI: Safe Evidential Real-World Reinforcement Learning Project URL: sevi-safe.github.io/ Paper: arxiv.org/abs/2610.31829v1

What if a single portrait could become a full-head 3D talking avatar driven by audio—without relying on parametric face models or per-instance optimization? AudioLift uses synchronized multi-view video generation as an intermediate step, followed by inconsistency-robust Gaussian reconstruction. This enables more expressive facial motion, supports details such as headwear, and recovers geometrically complete avatars with consistent appearance across the full 360° viewing range. Title: AudioLift: Audio-Driven Single Image to Full-Head Gaussian Avatar via Multi-view Video Generation Project URL: zhuoyuanjiang.github.io/audi… Paper: arxiv.org/abs/2610.31747v1

What if robot touch sensors could be designed for almost any shape, fabricated at scale, and transferred across platforms? OpenTac introduces an open ecosystem for customizable 3D tactile skins, combining a scalable resistive sensor with tools that generate fabrication files from 2D outlines, developable 3D surfaces, and even text-based descriptions. OpenTac decodes both normal force and local shear, while providing simulation, cross-sensor transfer, and adaptive data acquisition for control and learning. The authors designed and fabricated more than 30 sensors across robot embodiments, demonstrated tactile feedback and manipulation learning on four real-world robots, and achieved 19.2× faster partial readouts than full-array scanning on a 1024-taxel skin. Title: OpenTac: An Open Ecosystem for Customizable, Scalable, and Transferable 3D Tactile Sensing Project URL: opentac.github.io Paper: arxiv.org/abs/2610.31661v1

Reconstructing where a hand is in the real world—not just how it appears in video—is important for understanding human interactions and transferring them to robots. DewHand addresses this challenge from egocentric video with a unified model that jointly estimates hand and camera motion in a shared world coordinate system, helping avoid errors introduced by separate estimation stages. The method combines scene-level geometric features, metric-space supervision, and hand-specific features to recover scale and detailed articulation. Its flow-matching formulation also models plausible motion during occlusion or when hands leave the camera view. On HOT3D and ARCTIC, DewHand improves world-space accuracy and temporal consistency over prior methods, reducing rigid-aligned error from 32.9 to 28.4 mm and from 76.2 to 36.2 mm, respectively. Title: DewHand: Direct Metric World-Space Hand Motion Reconstruction from Egocentric Videos Project URL: narusuzuki.github.io/project… Paper: arxiv.org/abs/2610.29936v1

Robots may learn more effective bimanual dexterity by learning not only how hands move, but also the 3D space around them. SpatialDex reconstructs scene geometry and hand motion from RGB-only human videos, then uses both to pre-train a visuomotor policy for coordinated wrist and finger control. Pre-trained on 1 million human-video clips and fine-tuned with a small number of robot demonstrations, SpatialDex improves performance over prior methods by 16.5 points across nine simulated tasks and 38.3 points across six real-world bimanual tasks. In matched experiments, adding reconstructed geometry improves success by 14.2 points over hand-motion-only pre-training. Title: Reconstruction-Augmented Visuomotor Pre-Training of 3D-Grounded Bimanual Dexterity from RGB-Only Human Videos Project URL: spatial-dex.github.io/ Paper: arxiv.org/abs/2610.29716v1

Reliable robot planning requires more than recognizing a scene—it requires understanding which actions are possible and how they change the world. CausGround addresses this challenge by turning past interaction experience into reusable, action-centered causal knowledge about preconditions, effects, object attributes, and relationships. At planning time, it grounds task-relevant knowledge in current observations, reasoning backward from goals and forward from action effects to select executable multi-step plans—without updating model parameters. Across VirtualHome, RoboCasa, and SWITCH, CausGround improves planning and interaction reasoning, while real-world manipulation experiments demonstrate its ability to derive executable action sequences from MLLM-interpreted goals. Title: CausGround: Causal Interaction Grounding for Adaptive Embodied Planning Project URL: causground.github.io Paper: arxiv.org/abs/2610.29702v1

How can a model tell whether an object is made of rubber, metal, or something else when appearance alone is ambiguous? AutoPhyX introduces a text-conditioned approach for assigning spatially varying physical properties to 3D assets in a single forward pass. The framework combines semantically decomposed 3D training data with text descriptions and uses cross-modal modulation to connect language with visual features. Its voxel-field formulation supports meshes, point clouds, Gaussian Splatting, and NeRFs, offering a practical route toward more physically grounded simulation for embodied AI and robotics. Title: AutoPhyX: Automatic Text-Condition Physics Property Generation Project URL: hb-pencil-zero.github.io/Aut… Paper: arxiv.org/abs/2610.29517v1

Can robot policies learn from the failures and recoveries they encounter during deployment? TIDE (Training In DEployment) enables human-in-the-loop online post-training by learning directly from human interventions and policy-generated rollouts as a robot performs its tasks. Unlike online reinforcement learning methods that require a learned critic, reward model, or action scorer, TIDE uses provenance-conditioned supervised learning to update the policy during deployment. Across four real-world bimanual manipulation tasks, it improved task progress and full-task completion over the warm-start policy and matched additional offline demonstrations, while making updated checkpoints available during deployment. Each post-training session completed in under an hour, with learning and checkpoint handling running asynchronously. Title: Training in Deployment: Human-in-the-loop Online Post-Training for Robot Manipulation Project URL: logosroboticsgroup.github.io… Paper: arxiv.org/abs/2610.29264v1

How can a humanoid robot follow human motion in real time without sacrificing meaning—or physical feasibility? SMR addresses this challenge with a strictly causal, streaming neural retargeter. It learns from physics-based experts that account for contacts, dynamics, and actuator limits, then uses only current and past human observations to generate robot motion references for whole-body control. The result preserves motion semantics while adapting movements to the robot’s constraints, with improved closed-loop performance over kinematic and physics-refined baselines. Title: SMR: Learning streaming motion retargeting for humanoids from human semantic Project URL: logosroboticsgroup.github.io… Paper: arxiv.org/abs/2610.29169v1

A CAD assembly can look plausible—and still fail when asked to perform its intended function. FunCAD-Harness, a framework that turns abstract functional requirements into candidate-specific physical tests. It then uses simulation feedback to guide revisions. The authors also present FunCADBench, a benchmark covering 46 mechanism families with interdependent, multi-step operations. Experiments show that the approach reveals mechanical failures missed by visual inspection and improves functional success over standard simulation-augmented baselines. Title: Can It Actually Work? Synthesizing Functional CAD Mechanisms Project URL: thegenlab.ai/funcad/ Paper: arxiv.org/abs/2610.26861v1