Saturday Robotics @saturdayrobotic × IROS 2026 Research Night — part 2️⃣ of 4️⃣ Where the data comes from, and how you know it worked. @HarryXu12 (@PokeBot_Global/ 破壳机器人) — three bets for home robots
• Demo: autonomous Mapo Tofu on a dual-arm kitchen robot; publicly, a nine-minute cook with no human intervention.
• GPT-6 Astra as a scene factory: room photo → generated 3D scene → fine-tuning → real2sim replay.
• What's next: robot-free data > teleop; RL > DAgger; zero-shot > ICL. The robot-free wave: Sunday ACT-1 (zero teleop), Reward AI OM-1 (human data only), X Square TwinDEX (up to 5.3× teleop throughput), Generalist GEN-1 (500k+ hours from wearables).
• His timeline: usable robots within two years — not omnipotent, but well-defined products. Liangwang Ruan (@XSquareRobot) — X2Real
• Faithful: system-identified controllers keep sim and real arms within millimetres → 0.84 sim–real score correlation.
• Effective and diverse: 44 tasks, 10 capability dimensions, randomized over appearance, task setting and embodiment; 12,485 rigid + 670 articulated assets.
• Fair: a strict train/eval split plus ~300 h of sim trajectories.
• Evolvable: "Mana" — LLM agents write tasks in a YAML physical DSL, compiled into Isaac Lab-Arena.
• Best of four policies: 53.9% in-distribution → 34.3% out of distribution. 
1️⃣ If your home robot needs a demo, a prompt or a babysitter to finish dinner, it's a lab protocol.
2️⃣ Robot-free data and zero-shot are one bet: embodiment-agnostic data is what transfer is made of.
3️⃣ The Astra slide is the sleeper: once a photo becomes a sim scene, the bottleneck moves from assets to physics calibration.
4️⃣ If sim rankings don't move with real rankings, you have a render, not an evaluation.
5️⃣ Sim and real never match frame by frame, so scoring recovery measures what actually transfers.
6️⃣ 53.9% → 34.3% says more than any in-distribution leaderboard. #IROS2026 #HomeRobots #Simulation #RobotLearning #SaturdayRobotics

Welcome to Saturday Robotics @saturdayrobotic IROS Robotics Research Night!

When Will Robots Actually Be Ready for Our Homes? PokeBot CTO Zhengrong Xue believes robots could achieve zero-shot generalization across 80% of common household tasks within a year. But getting there may depend less on polished demos than on failed trajectories, sparse rewards, and better data. In this episode of Zhihu’s podcast 有限理性FM, we sit down with Huazhe Xu (@HarryXu12), a Tsinghua IIIS assistant professor and founder of @PokeBotGlobal, and his student Zhengrong Xue (@zhengrongX), now the company’s CTO. Their goal is ambitious: bring genuinely useful robots into ordinary homes. 🍳 Why make a robot cook mapo tofu? It is not simply a flashy demo. Cooking is a long-horizon, contact-rich task with dozens of opportunities for failure. PokeBot built its end-to-end demo in just over three months. The process exposed the real difficulty of household robotics: success requires generalizing across objects, layouts, environments, and unfamiliar combinations. 🧪 Why failed data may be more valuable A successful trajectory only shows what worked. A near miss reveals where the policy’s decision boundary actually lies. This is why disturbance tests and high-quality failure data matter. Repeating the same task in the same room for 100,000 hours may generate plenty of data, but very little new information. For household RL, the team also prefers sparse rewards: a clear 0 or 1 for final success, instead of dense human-designed signals that may lock the robot into one predefined path. 🧠 World models, VLA, and touch The conversation explores several open questions: 🔹 Should robots predict future visual states first, then infer the actions needed to reach them? 🔹 When does a world model offer something fundamentally different from a VLA? 🔹 Can elegant tactile systems justify their cost in real household deployment? 🔹 How can skills transfer across different robot bodies? Their central belief is that the architecture itself may not need to be extraordinarily complicated. The breakthrough may come from data that finally makes end-to-end learning work. 🏠 Why start with the home? Because homes are messy, unpredictable, and information-rich. That makes them difficult deployment environments, but potentially excellent places to train general-purpose embodied intelligence. Perhaps robotics will eventually have its own GPT-3 moment: not a system that solves every open problem individually, but one that makes many of those problems unnecessary to solve by hand. 🎧 Listen to the full conversation: xiaoyuzhoufm.com/episode/6a8… ▶ Watch on YouTube: youtu.be/O3E0rO7C1OE?si=TJ98… Follow 有限理性FM, a podcast by Zhihu, for more conversations with the researchers and builders bringing AI into the real world. #EmbodiedAI #Robotics #WorldModels #ReinforcementLearning #VLA #PhysicalAI

cool …the first robot to make Mapo Tofu in a single. Making tofu is actually quite a challenging Chinese culinary task… (Shanghai Pokebot is building home robots.)