| Welcome, sci-fi realistsā¦ | This week, the robot hand got a baseball tryout. | San Franciscoās Foundation developed a humanoid hand that caught a baseball midair with surprisingly human precision. Instead of cramming a motor into every finger, its ātendon-drivenā design moves the motors into the forearm and pulls the fingers with cable-like tendons. | That makes the fingers slimmer, lighter, and faster. Sensors at each joint track how far every finger bends, while the hand reshapes itself around whatever it touches. | For factories and warehouses, that could mean robots that handle awkward objects instead of one perfectly staged box. For Major League Baseball, spring training may eventually need a CAPTCHA. | Anyway, hereās what you need to know about robotics right now: | Google DeepMind gave one AI system control from feet to fingertips. NVIDIA opened a large reasoning model for autonomous driving. Tau, Sharpa, and Cosmic pushed robots toward real-world shifts. RL-100 reported 1,000 successful robot trials out of 1,000. U.S. regulators targeted foreign robot imports and robotaxi crashes. HumanCLAW exposed how poorly AI models understand their own bodies.
| | Google DeepMindās Gemini Robotics 2 controls humanoids from feet to fingertips | Google DeepMindās new Gemini Robotics 2 targets one of roboticsā hardest coordination problems. DeepMind wants one system to reason with the robotās whole body instead of treating walking, reaching, and gripping as separate tricks. | Gemini Robotics 2 is a vision-language-action model, or VLA. In plain English, it reads camera input and instructions, decides what should happen next, and turns that decision into robot movements. | Hereās what happened: | The model coordinates the robotās full body, from its feet to its fingertips. It can break a goal into several physical steps and adjust when the situation changes. DeepMind demonstrated it on Apptronik hardware performing navigation and delicate hand tasks.
| One example asks a humanoid to place a watering can inside a specific colored bin on a low shelf. The robot must identify the can, pick it up, walk to the shelf, bend down, and release it in the right place. | Why this matters: Most robot demos showcase one polished skill. Real work is a chain of skills, and one missed grip or bad step can ruin the entire job. A useful warehouse robot must move, balance, understand instructions, handle objects, and recover when reality refuses to follow the script. | Our read: The physical AI race is shifting from individual robot skills toward operating systems for entire bodies. DeepMind says field tests are next. The meaningful benchmark will not be one clean demo, but whether the same model can finish a full shift after people, objects, and lighting start changing around it. | | Robots in Action | DeepMindās video with Apptronik shows why the phrase āwhole bodyā matters. The robot walks to work areas, bends without toppling, transfers objects, and coordinates both hands without an obvious reset between every step. | The smaller movements are more revealing. One robot unscrews a lightbulb from a lamp. Another ties a knot in a trash bag, a task that requires the fingers to keep changing pressure and position. | Each action looks modest on its own. Chaining them together is the progress. The glamour shot is the lightbulb; the product test is doing it again after someone moves the lamp. | Introducing Gemini Robotics 2 |  | Gemini Robotics 2 brings whole body intelligence to robots |
|
| More Robots in the Wild |  | Tau Robotics' humanoid handles one of the least glamorous but most important deployment challenges: getting out of the van. |
| Tau Robotics is starting with the boring obstacles.Ā Tau showed a humanoid climbing out of a van, which sounds minor until you picture a service robot dealing with curbs, vehicle floors, doorways, and stairs. Oh, and Tau has also started an invite-only cleaning pilot in San Francisco. The robot still combines AI with human operation, while the company works toward reducing the human role to safety supervision. Light Origins taught one controller when to walk, climb, or vault.Ā LightParkour combines several movement skills inside one policy, the software that chooses a robotās next action. The robot decides when normal walking is no longer enough, performs the parkour move, and returns to walking. Transition training raised full-course success from 33% to 98% in simulation. Hardware videos show the policy transferring without additional real-world training, although the published numbers come from simulation. Satyress asked why a disaster robot should look completely human. Its threehalves concept combines a humanoid torso with a four-legged base, modular limbs, and quick-connect tools for fires, mines, industrial plants, rubble, and contaminated sites. The current concept is teleoperated, meaning a person controls it remotely, and Satyress has not shown a verified field deployment. Still, the design makes a useful point: the best body for dangerous work may look more like a mobile toolbox than a person. Cosmic Robotics taught one heavy machine a new construction job. In a company demonstration, Cosmic-1 detected, lifted, and installed large cooling pipes used in data centers. The team kept the same machine and software, swapped the end tool, and added task-specific training examples. Cosmic says the new behavior took about ten days to develop, including an overnight run using synthetic data, or computer-generated examples. Sharpa is putting a humanoid behind a real restaurant counter.Ā Sharpa North is scheduled to work a full shift at a Dairy Queen store in Shanghai this August, using the restaurantās existing tools and serving real orders. Each order takes 50 to 60 consecutive manipulation steps. The signature upside-down Blizzard flip also requires tactile feedback, or a sense of touch, so the robot can grip a flexible cup without crushing or dropping it. Enigma put 100 physical robots online to learn how people naturally direct machines. The company publicly launched with a $71M seed round and a browser experience that let people control real robots. Every instruction, correction, and confused attempt becomes data about how humans communicate with machines. Enigmaās larger bet is that one model and control system can work across many kinds of robot hardware.
| | Around the Horn | NVIDIA released Alpamayo 2 Super, an open autonomous-driving model with 32 billion parameters, a rough measure of its size. It can interpret a road scene, explain its reasoning, plan a route, and output driving actions. RL-100 learned from human demonstrations, then improved through trial and error using stored data and real robot practice. Under its predefined protocol, it completed 1,000 of 1,000 evaluated episodes across eight tasks and ran a mall juicing station for about seven hours without failure. CLIFT turned deployment failures into fresh training examples for Googleās closed Gemini Robotics On-Device model. After two cycles through Googleās hosted training service, the researchers reported near-perfect performance on several tasks that required precise physical contact. Zoox received temporary federal clearance for commercial deployment of its purpose-built robotaxis, which have no steering wheel or pedals. The company still needs the remaining state and local permissions before charging passengers. Andromeda Surgical completed New Zealandās medical-device registration for software that assists with minimally invasive urinary-tract surgery. The company describes New Zealand as its first commercial market; other regulatory clearances are still being pursued. The FCC blocked new approvals for foreign-made advanced robots, including humanoids and quadrupeds, over national-security concerns. Previously approved products are not covered by the new restriction. German robotics startup Agile Robots projected that it would double its revenue this year as manufacturers spend more on flexible automation. Reimagine Robotics went public with a pitch centered on robots that improve from experience instead of relying only on fixed programming.
| | Bots Behaving Badly | Now for the gap between the demo reel and the real world. | | Uber partner Avride is under investigation by the National Highway Traffic Safety Administration after 16 crashes in Austin and Dallas. Regulators said the vehicles changed lanes into nearby traffic, failed to slow for vehicles ahead, and struck objects partly blocking the road. | All 16 crashes occurred while a safety monitor was sitting in the driverās seat, and one involved a minor injury. NHTSA is now examining whether the problem reflects excessive assertiveness and insufficient driving capability. That is the autonomy industryās hardest question in miniature: can a system remain safe when an ordinary road stops behaving like its training data? | Vision-language models also keep forgetting they have bodies.Ā HumanCLAW tested nine models across 1,218 episodes that required finding an object, navigating toward it, and physically interacting with it. The best model completed the full sequence only 16.8% of the time. | The benchmark separates high-level decisions from low-level motor control, so a bad result cannot be blamed on weak joints or poor balance. Models recognized visible objects fairly well. They failed while exploring, tracking their own position, noticing collisions, and confirming they had reached the target. An AI can describe the chair in exquisite detail while forgetting where its knees are. | | Industry Signals | Full shifts are replacing single tricks. Sharpa, Tau, Cosmic, and Zoox are all testing whether robots can survive long sequences of work outside a controlled demo. Reliability is becoming the real leaderboard. RL-100 shows the upside of repeatable performance, while HumanCLAW shows how quickly embodied reasoning still falls apart. The robot stack is splitting into layers. General-purpose models interpret goals, task software chooses actions, and fast reflex-like controllers handle balance, grip, and touch. Deployment data is becoming a product. Enigma, CLIFT, Tau, and Cosmic all turn corrections, failures, and real-world attempts into better behavior. Policy is starting to shape the hardware market. Import restrictions, safety investigations, and medical approvals will determine which robots can be sold long before the technology feels finished.
| | | Hi! Iām Cameron Hashemi-Pour a technical writer and editor who covers AI, robotics, machine learning, Industry 4.0, cybersecurity, and emerging technology. Previously, I wrote for TechTarget/Informa TechTarget and contributed scripts for the āEye on Techā YouTube channel. | Thatās all for this week! | What'd you think of today's email? | |
|
| P.S. If you havenāt alreadyā¦ follow us on Youtube! |
|
