General Intelligence Labs

🤖 Saturday Robotics & World Model Reading Club #27.2. Rethinking Robot Development: Co-Designing Morphology, Sensing, and Learning @saturdayRobotic is happy to partner with @gi_labs to present: 🎤 @shumochu: Data Layer for the Next Billion Robots 👉🏻 xitter.gq/junfanzhu98/status/209… Robot foundation models don’t have an internet-scale action corpus. LLMs can scrape text; robot policies need real-world pose + vision + touch + hand motion, time-aligned and recorded from actual work. A million pouring videos ≠ synchronized end-effector pose, jaw state, contact force, or finger trajectories. @gi_labs: Scale with humans, not robots. Humans already perform the tasks—in kitchens, factories, construction sites, farms. Teleoperation adds a robot, rig, operator, and embodiment constraint: a throughput tax disguised as a data strategy. Capture the work where it happens. 🔥 Moat is bottom of stack: capture hardware→data operations→cleaning/labeling. Own layers 1–3 or rent your bottleneck. Sync is hardware or fiction; quality is a capture-time property. Hardware-clocked, factory-calibrated sensors turn pixels into metric trajectories; QA in the field prevents expensive post-hoc cleanup. 📹 EGO1-GS: stereo 2×1080p30 global-shutter cameras, 157°×84° FOV, 63-mm baseline, 400-Hz 9-DoF IMU, stereo audio, factory calibration, on-device hand detection + voice alerts; all streams share one clock and export to .mcap. Global shutter protects fast-motion contact frames; wide FOV protects data yield. 🦾 UMI0: bimanual handheld grippers, <1 mm XYZ, <0.1° orientation, <0.5 mm jaw, 5-MP 180° fisheye + jaw encoder, Quest tracking + on-device kinematics filtering. Bimanual capture in one take avoids phase error. Gen2: 535→450 g, 10→12 cm travel, dual forward/side global-shutter cameras, 157°×84°, while preserving accuracy. The second view attacks wrist-camera self-occlusion exactly at contact. 🖐️ Tactile: 77 Hall-effect sensing points/fingertip, 1-mm resolution, Fz 0–25 N, Fxy ±10 N, 10M-cycle life, synchronized with vision/pose/jaw. Touch is an action-space channel, for slip/contact geometry that RGB can't directly observe. 🧤 MoCap glove: full per-finger kinematics → multi-finger robot hands. One device cannot cover manipulation: parallel jaws, dexterous hands, head vision, wrist vision and tactile are different action spaces. Format fragmentation is a hidden scaling tax. 🚀 Proof: 506 handheld demos, 212,724 frames / 1h58m / 0 teleop→π0.5 fine-tuned on OpenArm achieved 18/20 (90%) real-robot pen insertions; tracking σ = 0.46/0.47/0.98 mm XYZ, 0.067°. The load-bearing zero is 0 teleop—but this is an existence proof, not “two hours solves general robotics.” 🔄 Capture→post-process→post-train→eval→capture. SLAM/VIO + keypoints → .mcap / LeRobot v3; failures decide what gets captured next. Progress = failure modes retired per cycle, not hours recorded. Vertically integrated capture is robotics’ equivalent of owning the corpus. The goal is a data layer that scales with people—not robots—for the next billion robots.

🚨 What if we’re designing robots the wrong way? @saturdayrobotic Saturday Robotics × @gi_labs General Intelligence 🤖 @shumochu will be joining us with a new announcement on EGO1GS and egocentric robotics capture. 👀 For Reading Club 27, we’re rethinking: Morphology × Sensing × Learning Instead of: Hardware → Sensors → AI What if we co-designed all three around the interactions we want robots to perform? Our keynote Chenyang Ma (@AppliedInt, @UNC) will explore: 🦾 Handroid — 27-DoF reconfigurable morphology ✋ Current as Touch — motor current as intrinsic tactile feedback 🧠 Learning designed around the body, not on top of it 📍 SF 📅 Sep 5 🕑 2–5 PM 🎟️ luma.com/jm9jceqr #Robotics #PhysicalAI #RobotLearning #EmbodiedAI