I was there Friday night in San Francisco for the @REK event, and this is something you had be in the room to get it, but I’ll try to describe it. It hit completely differently from robot-versus-robot fights. Those are entertaining, sure, but after a while they start to feel strange. You’re empathizing with machines that aren’t conscious, don’t feel pain, and don’t feel anything. Robot versus human hits different and raw, it was an emotional rollercoaster. I felt real fear for the human fighter, and real excitement, mostly fear, and when it ended there was a huge wave of relief. The robots were remote-controlled, of course, but the kicks were genuinely powerful. We know how power-dense those actuators are, and the human fighter took real hits to his body. The kicks from the six-foot-tall EngineAI T800 seemed nearly impossible to dodge when they were hurled in the right direction. The Unitree H1 was even harder to deal with because its dynamic balance is so good that it just refused to go down. This could be the start of a genuinely new entertainment genre. But I can tell you this: if someone wants to fight a robot with a real AI brain and real perception, you’d need robots that are far less powerful, throttled down, and compliant. Otherwise it ends in serious injury or worse. This is the beginning of something different in robotics and entertainment, and we need to think seriously about it before regulators use moments like this as fearmongering fodder to throttle the entire humanoid frontier.

Here’s what I’m watching tonight: 4 hours of Figure 03 robots doing home chores. Long, uncut scenes of housework in 30 different homes. Figure introduced Helix 2.5, its most advanced humanoid AI yet, built to answer one question: can a humanoid walk into a home it has never seen and get to work on its own? The setup One foundation model pretrained on Index, Figure’s global dataset of human behavior, then fine-tuned into three whole-body behaviors: tidying living rooms, folding towels, and making beds. Tested across 30 Bay Area homes with zero data collected in any of them, and with objects it had never seen. Results • 56% zero-shot success vs. 9% trained from scratch. Index pretraining was the only variable. • No partial credit in success rate. Every toy in the basket, every towel folded, the whole bed made. • Half the task data of a comparable Helix 02 behavior, and 30× wider generalization. • A human-to-robot transfer scaling law: doubling Index data improved action prediction predictably enough to forecast the largest run’s loss to four decimals. • Whole-body self-correction. The robot steps back, repositions, and walks around the bed to fix a fold. Index now generates ~35 minutes of new human experience every second, and Figure has committed $3.5B of compute to Helix.

Agility showcased the bipedal Digit 5 in the launch video, but also gave glimpses of a wheeled form working at a grocery store and a datacenter. Agility's target applications have been logistics and warehousing, where floors are flat and moving space isn't really a constraint. A wheeled base makes more sense there. They've announced $300M+ in multi-year customer orders for Digit 5. How many of those will be served by the bimanual wheeled form? My guess: the vast majority.

We benchmarked only 7 hours of Human Archive data against 4 widely used public datasets totaling 17 hours: HOT3D, H2O, HOI4D, and TACO. The results: Human Archive only: 85% task success | 65.6% task progress 4 public datasets: 57.5% task success | 52.5% task progress Mixture of all 5: 60% task success | 56.3% task progress

A humanoid can nail the grasp and still fail the task. Opening a low drawer isn't just the hand. It's the stance, the weight shift, and staying balanced through the whole reach. Hand trajectories alone don't capture any of that. That's what makes @GenrobotAI's whole-body mesh approach interesting. Using vision alone, they reconstruct full-body motion from first-person video, reporting ~3 cm whole-body error (about 2 cm upper body, 3.5 cm lower). The hard part: an ego camera barely sees the body. Hands vanish behind objects, arms cross, fast motion blurs, and long tasks drift. Nailing one clean frame is easy. Keeping a whole task consistent over time is the real problem. And they go past the mesh itself. The pipeline aligns hand-object contact (when it starts, holds, and releases) and runs automated quality checks to output model-ready data. A robot doesn't just need a pose. It needs how motion, contact, and object interaction unfold together. Their reported capacity: 100K hours of this a month. The story is the combination: detailed reconstruction plus a scalable pipeline to produce usable data. As humanoids move past the tabletop, whole-body data could become a core part of the next training-data paradigm.

Agility is moving away from its ostrich-like, backward-bending knee design (essentially a high ankle joint) toward more human-like legs with its new-generation humanoid. Digit 5 has just been unveiled. - New cycloidal hip/knee actuators, 50 lb payload, 20-hour shifts, 10:1 run-to-charge (90-min battery, 9-min charge). - It stands 5'11" (1.81 m), weighs 284 lb (129 kg), and reaches up to 7.2 ft (up from Digit 4's 5.5 ft). - Agility claims $300M+ in multi-year customer orders for Digit 5. Early access deliveries start in H1 2027, and they'll also be expanding into the EU and UK. This lands as Agility is set to go public via a $2.5B SPAC merger with Michael Klein's Churchill Capital Corp XI (ticker AGLT), expected to close later in 2026 pending shareholder and SEC approval.

Reward AI is based in the SF Bay Area. The founders are two of Stanford's standout robot-learning researchers: Zipeng Fu (Cofounder, CEO), advised by Chelsea Finn, ex-Google DeepMind. Lead author on Mobile ALOHA (the viral $32k robot that cooks and cleans) and HumanPlus (humanoids learning by shadowing humans). Chen Wang (Cofounder, CTO), advised by Fei-Fei Li and Karen Liu. Lead author on DexCap, the portable dexterous-manipulation capture system OM-1's wearable hand is built on.

Reward AI just came out of stealth and unveiled OM-1, a robot policy learned entirely from humans. No teleoperation, no on-robot data. One Model, Any Body. 1. Capture. A wearable 7-DoF hand (Omnibody Hand) lets people demonstrate manipulation naturally while working, cooking, or sorting. Sensors record tactile, proximity, force and hand pose at full human speed. 2. Learn. OM-1 trains on this human data alone. No teleop, no on-robot experience. Then it runs across bodies from industrial arms to humanoids, and scaling is just adding more human data. 3. Control. A high-frequency control layer, trained with RL in sim, executes those actions on any robot, holding steady through unknown loads like a stuck fridge door. The result: OM-1 picks up a brand-new task from under 30 minutes of data.

Skild's founders just dunked on the entire demo-industrial complex. Skild just disclosed a $100M annual revenue run rate, only 10 months after their first commercial deployment. 60+ paying customers, including G10 Fulfillment, Mitsui & Co, and STN Inc (data center infrastructure), with the vast majority of revenue coming from manipulation-heavy work (mobility is only 10%, Fetch solutions 4%). They argue that cherry-picked, successful demos make it feel like a use case is solved, but squeezing out the last 5% of reliability at acceptable throughput is the hard, under-appreciated part. The R&D is hard enough on its own. But real-world deployment also takes a mountain of dirty work behind the scenes: installation, integration, repairs, adapting to new processes and more. Skild plans to crack robotics RSI (recursive self-improvement). The flywheel: start with one general foundation model, deploy it into many different applications where in-context learning lets it adapt on the fly, then feed the data from all those specialized deployments back into the base model. Experience that adds a little bit to a specialist is gold to a generalist learning many tasks. Each cycle, the next deployment starts from a stronger base model and needs less specialization to get to work. The era of deployment begins.

Don't sleep on Tesla in humanoids just because they've gone quiet. Tesla bet on solving driving intelligence from vision alone, put a powerful inference computer in every car long before FSD was solved or profitable, and built the biggest robotics data flywheel. General-purpose humanoid needs the same systems, but on steroids. Debate the AI model architectures and actuator designs all you want. They definitely matter for setting the right direction. But what ultimately decides the outcome is the organizational strength that powers every decision. The trifecta: • Engineering: fast iteration, bringing new work in-house, pivoting quickly • Physical ops: manufacturing, data collection, supply chain, repairs, fleet management • Inference and training compute infra at massive scale Few companies can bring all three together at scale as well as Tesla.

Training general-purpose humanoids will need unprecedented compute capacity, and Figure is stacking the chips. Figure, with cloud partner Nscale, is deploying up to 100,000 NVIDIA Vera Rubin GPUs. It's committing $3.5 billion initially, with plans to scale to $6+ billion. Deployment starts in the second half of 2027, at a data center in Barstow, Texas.

Nvidia is now a humanoid robot designer. That's the real takeaway from this news. Nvidia is acquiring Hugging Face for $12.93B. Hugging Face is the largest open-model and dataset hub, the "GitHub of AI," with 3M+ models and 18M+ developers, robotics included. But here's the interesting part: Hugging Face has a serious robotics team from its acquisition of Pollen Robotics, which designs and builds open-source hardware platforms for robots, including humanoids (Reachy). Pollen team now gets the backing of the world's most valuable company. $NVDA