{"slug": "robowild-benchmark-claude-fable-5-2-leads-at-13-3-while-robodojo-hits-64-9", "title": "RoboWild Benchmark: Claude Fable 5.2 Leads at 13.3% While RoboDojo Hits 64.9%", "summary": "Claude Fable 5.2 leads XLANG Lab's public RoboWild leaderboard at 13.3% weighted success across 341 real-world-style robot tasks, ahead of GPT-6 Astra at 11.6% and Gemini 4 Carbon at 10.6%, while the same model family scores 64.9% on RoboDojo Sim in Anthropic's November 9 launch table. RoboWild spans 38 furnished scenes and 4,480 objects across 23 scene types at 3.60 objects per square meter, and grasp failure accounts for 42.6% of failed runs, with 87% of runs disturbing unrelated objects. The benchmark from XLANG Lab, the group behind OSWorld-style computer-use stacks, is positioned as the next frontier after digital workflows, with the dataset released on Hugging Face at roughly 18.5 GB under Apache-2.0.", "body_md": "**Claude Fable 5.2 tops the public [RoboWild](https://robowild.xlang.ai/?ref=explainx) leaderboard at 13.3% weighted success**, yet that is still only about **one task in eight** across **341** real-world-style robot chores. The same model family scores **64.9% on RoboDojo Sim** in [Anthropic's November 9 launch table](https://explainx.ai/blog/claude-fable-5-2-mythos-5-2-launch-pricing-api-2026). **RoboWild is the gap made visible**: dense rooms, articulated objects, and safety side effects that simulation stacks smoothed away.\n\nThe benchmark comes from **XLANG Lab**, the group behind **[OSWorld-style computer-use stacks](https://explainx.ai/blog/alibaba-opengcua-open-stack-generalist-computer-use-agents-osworld-v2-2026)**. Two years after OSWorld, frontier models handle many **digital** workflows; **physical robot-use agents (RUAs)** remain the **next frontier**, in the lab's wording on the site.\n\n## TL;DR: scores and what they mean\n\n| Question | Answer | \n|---|---|\n| Best model on RoboWild? | Claude Fable 5.2 at **13.3%** (ReAct, highest thinking) | \n| Runner-up? | GPT-6 Astra **11.6%** ; Gemini 4 Carbon**10.6%** | \n| Tasks / scenes? | **341** tasks,**38** furnished scenes,**4,480** objects | \n| Robots? | Mobile ALOHA, Franka Mobile, Unitree G1, Dexmate Vega-1 | \n| Top failure mode? | **Grasp failure** 42.6% of failed runs | \n| Sim comparison? | Opus 5.5 **50%** RoboDojo vs**4.81%** RoboWild (100 calls) | \n| Assets? | [Hugging Face dataset](https://huggingface.co/datasets/xlangai/RoboWild?ref=explainx) (~18.5 GB, Apache-2.0) | \n\n## What RoboWild adds beyond RoboDojo and BEHAVIOR\n\nRoboWild home entryway furnished scene for robot-use tasks\n*Image: [XLANG Lab](https://robowild.xlang.ai/?ref=explainx)*\n\n[RoboWild's site](https://robowild.xlang.ai/?ref=explainx) argues existing robotics benchmarks were **not demanding enough**: small workspaces, simple object sets, and **narrow task distributions**. RoboWild targets **wild** robot use: **homes, workshops, laboratories, and stores** with **3.60 objects per square meter**, which the authors report is **nearly twice** the next densest benchmark in their comparison table (**Gym-Anything** at 1.82).\n\nCoverage numbers from the official page:\n\n- **23 scene types** versus single-scene or low-diversity suites.\n- **61.9%** of tasks need**at least two** of eight measured skill types.\n- **65.7%** require**dependency chains** of two or more subgoals.\n- **58.6%** need**visual disambiguation** among similar objects.\n\nTask examples on the demo page include mundane but hard chores: **stack mail in a drawer**, **hang umbrellas on specific hooks**, **fold glasses onto a plate**, **open a locked door**. Those are the tasks humans dismiss as \"easy\" until a robot must do them with **three wrist cameras** and a **JSON tool interface**.\n\nIf you tracked **[Waddle-1 on RoboDojo-RC](https://explainx.ai/blog/waddle-1-robodojo-rc-81-score-claimed-vs-verified-november-2026)** this week, read RoboWild as the **opposite stress test**: not **six real bimanual tasks** on a leaderboard row, but **hundreds of diverse scenes** where **87% of runs disturb unrelated objects**, per the site's safety statistics.\n\n## Leaderboard: Fable 5.2 first, everyone still low\n\nThe published **ReAct** leaderboard on robowild.xlang.ai (five models, highest thinking per model) ranks:\n\n| Rank | Model | Weighted success | \n|---|---|---|\n| 1 | Claude Fable 5.2 | **13.3%** | \n| 2 | GPT-6 Astra | 11.6% | \n| 3 | Gemini 4 Carbon | 10.6% | \n| 4 | Claude Opus 5.5 | 6.5% | \n| 5 | GPT-6.1 Sol | 4.6% | \n\nThe abstract on the same page notes **token economics**: **Fable 5.2, Opus 5.5, and Carbon** consume **several times more output tokens** than **GPT-6.1 Sol and GPT-6 Astra** per task. Routing logic that picks **cheap models** for coding may need a **different policy** for **multi-minute robot episodes**.\n\nAnthropic's launch post already showed **Fable 5.2 at 64.9% on RoboDojo Sim** versus **39.4% for Opus 5.5** and **27.6% for Fable 5.1**. RoboWild **preserves the ordering** (Fable above Opus) but **crushes absolute success**: the site plots **Opus 5.5 at 50.00% on RoboDojo and 4.81% on RoboWild** with **100 tool calls**.\n\nThat is the headline for **[top Fable 5.2 demos](https://explainx.ai/blog/top-20-things-built-with-claude-fable-5-2-matt-shumer-anthropic-mollick-2026)** too: impressive **puzzle cubes and bottle caps** in curated clips are **not** the same distribution as **341 wild tasks** with **grasp failures** dominating.\n\n## Task design: perceptual and physical skill mix\n\nRoboWild publishes a skill requirement breakdown on robowild.xlang.ai that explains why **13% success** is not a formatting error.\n\n**Perceptual demands** (share of tasks requiring the skill):\n\n- **Visual disambiguation** : 58.6% — picking the right mug, hook, or letter among lookalikes.\n- **Visual reasoning** : 14.6% — layouts and spatial relations beyond OCR.\n- **Text reading (OCR)** : 13.4% — addresses on mail, labels on controls.\n- **Memory** : 5.0% — carrying state across subgoals.\n\n**Physical demands**:\n\n- **Articulation or material interaction** : 53.6% — doors, dials, deformable items.\n- **Precise manipulation** : 48.5% — stacking, folding, placement tolerances.\n- **Navigation** : 28.5% — moving bases through clutter.\n- **Bimanual manipulation** : 21.3% — two-arm coordination.\n- **Physical tool use** : 12.6% — drills, stand mixers, vise-like objects listed in the interactive object gallery.\n\n**61.9%** of tasks require **at least two** skill types; **0.8%** require five or more. That composition is why **computer-use agents** trained on **flat screenshots** do not transfer cleanly: the observation stream is **multi-camera**, **temporally extended**, and **physically consequential**.\n\n## Harness: how a VLM actually drives the robot\n\nRoboWild documents a **ReAct-style** loop on robowild.xlang.ai:\n\n1. The model receives a **task instruction** ,**robot-use guidance** (frames, units, quaternions),**tool schema** , and**three images** (head, left wrist, right wrist).\n2. It calls **`move_robot`** (and**` done`** ) — one tool per turn.\n3. The simulator advances and returns **observation images** until termination.\n\nEmbodiments share that interface across **mobile dual-arm**, **Panda on a base**, **humanoid**, and **wheeled humanoid** platforms. The benchmark is explicitly about **cross-embodiment generalization**, not a single arm in an empty bench.\n\nAblations mentioned on the site (not fully tabulated on the landing page) include **adding depth** and **enabling code execution** as ways to **raise success**. Builders replicating OSWorld-style **computer-use** stacks should assume **robot-use** needs **similar harness work** plus **physics-side safety**.\n\n## Failure modes: grasp, side effects, and humanoid pain\n\nRoboWild's results section is unusually blunt for a leaderboard launch:\n\n- **Grasp failure** :**42.6%** of the**188 failed runs** analyzed.\n- **Motion or execution failure** :**28.7%** .\n- **Functional object operation failure** :**16.0%** .\n- **Goal or target selection error** :**6.9%** .\n\n**Failure recovery** rates differ by error class: **Interface** errors recover locally **78.7%** of the time; **Contact** errors only **25.7%**. Agents can **retry UI-like mistakes** more often than **physical slips**.\n\nThe **safety and side effects** block is what product teams should quote:\n\n- **Unrelated objects disturbed** :**87%** of runs.\n- **Collision stops** :**78%** .\n- **Task objects dropped or tipped** :**75%** .\n\nThat aligns with regulatory attention on **robotaxi perception** after **[California's Waymo freeway order](https://explainx.ai/blog/waymo-california-dmv-freeway-ban-november-2026)**: low success is not only **task failure** but **collateral motion** in cluttered spaces.\n\nThe site also states VLMs **struggle on the humanoid embodiment** specifically, while **depth** and **code execution** help in separate harness tests. If you evaluate **[Perceptron Isaac checkpoints](https://explainx.ai/blog/perceptron-isaac-0-5-ready-to-run-checkpoints-yam-lerobot-november-2026)** for **VLA policies**, RoboWild is the complementary question: **what happens when a general VLM steers the same hardware through a tool API?**\n\nRoboWild embodied AI evaluation benchmark comparing sim scores to wild robot tasks\n\n## How to try RoboWild assets and code\n\nPrimary sources:\n\n- **Site and leaderboard** :[robowild.xlang.ai](https://robowild.xlang.ai/?ref=explainx)\n- **Paper PDF** : linked from the site citation to Hugging Face (`RoboWild.pdf` on the dataset repo)\n- **GitHub** :[xlang-ai/RoboWild](https://github.com/xlang-ai/RoboWild?ref=explainx) (installation guide referenced from HF)\n- **Dataset** :[xlangai/RoboWild](https://huggingface.co/datasets/xlangai/RoboWild?ref=explainx) —**18.5 GB** OpenUSD scenes, objects, robots, materials for**Isaac Lab**\n\nQuick start from the dataset card:\n\n```\ngit lfs install\ngit clone https://huggingface.co/datasets/xlangai/RoboWild robowild_assets\nexport ASSETS_DIR=\"$PWD/robowild_assets\"\n```\n\nThen follow the **RoboWild installation guide** on GitHub for simulator setup. The repo layout lists **scene01–scene38**, reusable **objects/**, **robots/** for all four embodiments, and shared **materials/**.\n\nThere is **no substantive Hacker News thread** yet under the RoboWild name (search checked November 21, 2026). For developer reactions on **robot evals**, see **[uni-robot-agent harness effects](https://explainx.ai/blog/uni-robot-agent-open-qwen-robodojo)**.\n\n## Benchmark landscape: density and scene breadth\n\nRoboWild's comparison table on the official site positions it against **RoboCasa365**, **BEHAVIOR-1K**, **Embodied First Exam**, **LIBERO**, **RoboARCade**, and **RoboDojo** on **scene types** and **objects per square meter**. RoboWild reports **23 scene types** at **3.60 objects/m²**; **RoboCasa365** shows **1 scene type** at **1.41 objects/m²** in the same table.\n\nThat is not an attack on prior benchmarks — each suite serves a purpose. **RoboDojo** remains the **simulation slice** Anthropic uses in launch materials and **RoboCurve** adapts for **real arms**. RoboWild is the **stress test for generalist VLMs** that must **pick tools, disambiguate clutter, and avoid side effects** in furnished spaces.\n\nTeams comparing **[AA-Robotics humanoid scores](https://explainx.ai/blog/aa-robotics-v0-1-frontier-models-robot-arm-humanoid-benchmark-claimed-vs-verified-2026)** or **[Gemini Robotics factory demos](https://explainx.ai/blog/gemini-robotics-2-5-mercedes-kitting-asimov-jailbreak-2026)** should add RoboWild when the question is **\"can a frontier chat model operate my robot API?\"** rather than **\"does a specialized VLA policy track boxes?\"**\n\n## What to do with these numbers in production\n\nPractical takeaways if you ship **robot-use** or **VLA** features:\n\n1. **Treat sub-15% success** as the**current ceiling** for general VLMs on**wild** tasks, not an outlier.\n2. **Budget tokens and wall clock** per episode; Fable-class models**spend more output tokens** per task on the official abstract.\n3. **Log side effects** , not only task success —**87% disturbance rate** is a**liability** metric.\n4. **Compare harnesses** , not raw model names: RoboDojo, RoboDojo-RC, and RoboWild use**different scenes, caps, and rubrics** .\n5. **Pair sim training** with**furnished-scene eval** before marketing \"general robot intelligence.\"\n\nFor **[Fable versus Opus routing](https://explainx.ai/blog/claude-fable-5-2-vs-opus-5-5-benchmarks-cost-routing-which-model-2026)**, RoboWild reinforces **Fable for long-horizon physical tool loops** — but **only if** your product tolerates **cost per attempt** and **low absolute success** today.\n\nFrontier comparisons that include **Gemini 4 Neon** and **GPT-6 Astra** live in our **[November frontier table](https://explainx.ai/blog/gemini-4-neon-vs-claude-opus-5-5-vs-gpt-6-astra-vs-carbon-comparison-november-2026)**; RoboWild adds **robot-use** as a row you should **not** infer from **coding benchmarks** alone.\n\n## Related on explainx.ai\n\n*Leaderboard figures and failure statistics are accurate as of November 21, 2026 per [robowild.xlang.ai](https://robowild.xlang.ai/?ref=explainx) and [Anthropic's Fable 5.2 announcement](https://www.anthropic.com/claude-fable-and-mythos-5-2?ref=explainx).*", "url": "https://wpnews.pro/news/robowild-benchmark-claude-fable-5-2-leads-at-13-3-while-robodojo-hits-64-9", "canonical_source": "https://explainx.ai/blog/robowild-fable-5-2-13-percent-xlang-robot-benchmark-november-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog", "published_at": "2026-11-21 00:00:00+00:00", "updated_at": "2026-11-21 05:12:14.952817+00:00", "lang": "en", "topics": ["robotics", "ai-research", "ai-agents", "ai-safety"], "entities": ["XLANG Lab", "Claude Fable 5.2", "Anthropic", "GPT-6 Astra", "Gemini 4 Carbon", "Claude Opus 5.5", "RoboWild", "RoboDojo Sim"], "also_reported_by": [], "takeaway": "Frontier models top out at 13.3% weighted success on RoboWild's 341 physical tasks versus 64.9% on RoboDojo Sim, and grasp failure causes 42.6% of failed runs, so simulation scores overstate real-world robot-use capability.", "recommendation": {"advice": "Evaluate robot-use agents on dense real-world-style scenes like RoboWild's 341 tasks rather than simulation alone; the same model family dropped from 64.9% on RoboDojo Sim to 13.3% on RoboWild.", "playbook": "evals-testing", "playbook_url": "https://wpnews.pro/playbooks/evals-testing"}, "alternates": {"html": "https://wpnews.pro/news/robowild-benchmark-claude-fable-5-2-leads-at-13-3-while-robodojo-hits-64-9", "markdown": "https://wpnews.pro/news/robowild-benchmark-claude-fable-5-2-leads-at-13-3-while-robodojo-hits-64-9.md", "text": "https://wpnews.pro/news/robowild-benchmark-claude-fable-5-2-leads-at-13-3-while-robodojo-hits-64-9.txt", "jsonld": "https://wpnews.pro/news/robowild-benchmark-claude-fable-5-2-leads-at-13-3-while-robodojo-hits-64-9.jsonld"}}