エピソード

  • HexaShell: Can Debugging Code Generalize RGB Robot Manipulation?

    HexaShell: Can Debugging Code Generalize RGB Robot Manipulation?
    2026/10/19
    Can a robot improve at manipulation by debugging code instead of updating neural-network weights? This episode explains how HexaShell turns robot experience into reusable tools, and what its RoboDojo score actually establishes. We cover three questions: how calibrated RGB observations become metric positions; why the reported 39.7% success rate uses category-balanced scoring; and how holding out layouts differs from holding out tasks. The experiments use a simulated dual-arm ARX X5 with head and wrist cameras, camera calibration, and proprioception. Along the way, we connect the design to Code as Policies and URAI, examine an inspectable geometric helper, and separate tool development from execution with frozen tools. The central limitation is a mismatch in task-level learning conditions: HexaShell develops tools on tasks in a category that RoboDojo's published training split reserves for transfer. For researchers building language-driven manipulation systems, the episode offers a concrete way to think about code as learned policy state. If software is part of what a robot learns, where should the training boundary be drawn — and what would convincing generalization look like?
    続きを読む 一部表示
    12 分
  • Niantic Spatial Gaussian Splats: Better Sim-to-Real Robot Training?

    Niantic Spatial Gaussian Splats: Better Sim-to-Real Robot Training?
    2026/10/19
    Niantic Spatial's Places Library is the starting point for this episode of Embodied AI 101, which examines how Gaussian splatting, aligned collision meshes, and existing simulation tools fit together, why visual realism and physical correctness must be assessed separately, and what those differences mean for teams building robot navigation systems for real warehouses and other deployment sites. We unpack the evidence behind reconstruction quality and sim-to-real transfer, examine the reported office-navigation comparisons, and turn the technical details into a practitioner's checklist covering downloadable assets, licensing, reproducibility, simulator integration, mesh defects, lighting variation, and the engineering work required to turn a captured place into a useful policy-training environment. Rather than treating an attractive demo as proof of general-purpose autonomy, we ask how to compare this workflow fairly with synthetic environments, where site-specific specialization might be valuable, and which measurements would justify adding it to a production training stack—so what would a convincing held-out warehouse evaluation actually need to demonstrate?
    続きを読む 一部表示
    12 分
  • Gemini 4 Argon for Robotics: What the Benchmarks Actually Show

    Gemini 4 Argon for Robotics: What the Benchmarks Actually Show
    2026/10/19
    This episode of Embodied AI 101 examines Google's Gemini 4 Argon robotics announcement, separating the demonstrated workflows from the broader promise by tracing how a calibrated camera view becomes a numerical motion command, why an ERQA score is not a manipulation success rate, and where a hosted reasoning model might fit alongside an existing robot controller and its safety mechanisms. We examine the reported sim-to-real improvement in a downstream VLA, the difference between a visually convincing digital twin and a model that predicts contact, the evidence practitioners can inspect or rerun, and independent Robocurve measurements that put latency and partial-credit scoring at the center of the deployment discussion rather than treating stronger reasoning as sufficient evidence of useful physical performance. For roboticists considering a pilot, which experiments, safety boundaries, held-out tasks, and reproducibility artifacts would justify trusting this approach beyond a curated demonstration—and could Argon's biggest near-term contribution be helping engineers build the simulations and training data behind the robot rather than controlling it directly?
    続きを読む 一部表示
    12 分
  • GreenWAM: Can One World-Action Model Learn Across Robot Bodies?

    GreenWAM: Can One World-Action Model Learn Across Robot Bodies?
    2026/10/18
    Can a robot foundation model learn a reusable connection between body motion and the scenes that motion creates? This episode examines GreenWAM's architecture, controlled experiments, and open benchmark release to clarify what joint world-and-action learning contributes to manipulation. We cover three questions: how anatomical coordinates organize experience from different robots; why retaining extra prediction objectives improved randomized RoboTwin success; and what the LIBERO-plus result says about robustness to changed observations and task conditions. The discussion connects GreenWAM to Cosmos 3, Unified World Models, and RDT-1B, separating the inherited ideas from Sber's robotics-specific training recipe. The evaluation spans simulated manipulation benchmarks and a physical packing task on the Green humanoid. The main limitation is the absence of a clean experiment isolating transfer to a genuinely unseen body from target-specific adaptation. For researchers building generative robot controllers, the practical question is whether world prediction should remain an active training objective as the policy specializes. What changes when a controller must learn both what to do and what its actions will make happen?
    続きを読む 一部表示
    13 分
  • 3DCodeVerse: Can Code Generation Build Better Worlds for Robotics?

    3DCodeVerse: Can Code Generation Build Better Worlds for Robotics?
    2026/10/18
    Can code-generating agents make 3D worlds easier to build, edit, and use in robotics? This episode explains 3DCodeVerse's training corpus, agent harness, and open-model experiments to show where its improvements come from. We cover three findings: how roughly two million code–condition pairs relate to the underlying program collection; why execution, geometric measurements, and visual feedback belong in the same repair loop; and what the reported human-preference and execution-success results actually measure. The discussion follows static objects, articulated mechanisms, and scenes through Blender, CAD tooling, browser graphics, and URDF, then examines the paper's MuJoCo microwave-interaction experiment. We connect the approach to Infinigen and SceneCraft and discuss its relevance to simulation infrastructure and code-model training. The central limitation is the robotics validation gap: preserving door-and-handle geometry does not establish correct interaction or improved policy learning, and even a motionless image passes the reported contour criterion. What would it take to turn an editable generated scene into a controlled, useful robot-learning experiment?
    続きを読む 一部表示
    14 分
  • Motion- and Gravity-Aware 3D Reconstruction: Do Offline Priors Help?

    Motion- and Gravity-Aware 3D Reconstruction: Do Offline Priors Help?
    2026/10/17
    Can optical flow and gravity improve camera-based 3D reconstruction without becoming permanent parts of the inference pipeline? This episode explains the training design, benchmark evidence, and practical implications of motion- and gravity-aware feed-forward reconstruction. We cover three questions: how training-only geometric hints enter VGGT and Pi3, why residual flow is a useful alternative to full optical flow, and what architecture-dependent injection points reveal about geometric representations. The discussion connects these choices to privileged-information learning and to the surrounding ecosystem of pretrained geometry tools. It focuses on multi-view reconstruction using real ETH3D imagery and synthetic TartanAir wide-baseline scenes, with robot perception as the intended application context. The central limitation is whether the evaluation actually isolates dynamic-scene robustness: difficult viewpoint changes and improved average geometry do not automatically establish better handling of independently moving objects. TartanAir's wide-baseline benchmark is explicitly biased toward static objects. For researchers choosing between a larger deployed model and a richer training pipeline, where should the next unit of engineering effort go?
    続きを読む 一部表示
    13 分
  • ASPIRE: Can Code Refinement Make Robot Manipulation Generalize?

    ASPIRE: Can Code Refinement Make Robot Manipulation Generalize?
    2026/10/17
    Can a robot improve by remembering how it fixed its own code? This episode explains how NVIDIA's ASPIRE turns robot execution feedback into reusable manipulation skills, and what its evaluations actually establish. We cover three questions: how fine-grained traces support targeted program repair; why a persistent skill library changes the meaning of continual learning; and how task-specific refinement differs from transferring experience to unfamiliar problems. The analysis separates the June 2026 paper from the October 2026 open-source update and its reported near-saturated benchmark results. ([github.com](https://github.com/NVlabs/ASPIRE)) Experiments span LIBERO-Pro and Robosuite simulation, household mobile manipulation, and skill transfer to a physical bimanual YAM robot. ([arxiv.org](https://arxiv.org/html/2607.00272v1)) We connect the method to Voyager, CaP-X, and the emerging combination of coding agents with learned motor primitives. The central limitation is evaluation scope: success after debugging a known task does not establish broad competence on unseen physical task families. The released BEHAVIOR workflow makes that distinction especially concrete. ([raw.githubusercontent.com](https://raw.githubusercontent.com/NVlabs/ASPIRE/main/.claude/skills/iterative-debugging/behavior/SKILL.md)) For researchers building robot learning systems, the practical question is compelling: what should a failed execution teach the next program?
    続きを読む 一部表示
    13 分
  • UAV-ESPrune: Can Visual Token Pruning Preserve UAV Reasoning?

    UAV-ESPrune: Can Visual Token Pruning Preserve UAV Reasoning?
    2026/10/16
    Can a UAV vision-language model discard most of its visual tokens while preserving the evidence needed for counting and planning? This episode explains UAV-ESPrune's selection strategy, how to interpret its results, and the comparison that most affects its scientific claim. We cover three practical questions: how protected local evidence differs from generic feature diversity; what accuracy retention and end-to-end timing tell you about different parts of an inference system; and what makes a pruning baseline a faithful comparison. The setting is Qwen2.5-VL-7B inference on stored real UAV imagery, timed on an RTX 4090. We connect the design to FastV, VisPruner, and the question-guided mechanisms of published SparseVLM. The central limitation is the simplified SparseVLM adapter used in the comparison. For researchers working on aerial perception and efficient multimodal inference, the episode offers a task-centered way to think about compression: which distinctions must remain recoverable for an answer to stay correct? When two objects look alike, should your compressor treat them as redundant—or as separate evidence?
    続きを読む 一部表示
    13 分