A Humanoid Cleans an Apartment for $30 an Hour, Six AgiBot Robots Run a Store for 24 Hours

A humanoid robot cleaned a reporter's apartment this week for $30 an hour, and a person at the company's headquarters was steering it the whole time. Six AgiBot humanoids ran a convenience store in Shanghai through a 24-hour livestream and finished 1,730 tasks, with nothing published about who helped them off camera. And a group of 28 researchers took one open model, changed none of its weights, and watched its score on a two-armed robot benchmark move from 1.9 out of 100 to 31.4 by swapping the software wrapped around it.

Tau Robotics sends a piloted humanoid to clean an apartment

Tau Robotics, a San Francisco startup, sent a roughly 4.5-foot humanoid with a 3D-printed head and two claw hands to clean a reporter's apartment for MIT Technology Review. People pilot it remotely from the company's headquarters through cameras on its head and wrists, and bed-making and laundry run on kinematic guidance, where the robot copies the pilot's own arm movements rather than executing a learned policy. The reporter's comparison was a radio-controlled car more than a Waymo. It cost $30 for the hour.

The failures are the most useful part of the piece. The first visit ended early when the remote-control equipment overheated and the robot stalled. It had to be carried up the carpeted stairs. The toilet-cleaning wand head fell off three times, and the robot screwed it back on each time, which is finer manipulation than cleaning the toilet asked for. Cofounder Cornelia Weinzierl says it can already vacuum and mop unaided, and that "the amount of work and time that we spend just optimizing, for example, timings and connectivity is crazy." Tau says it has completed a couple of hundred cleanings.

A related business runs at warehouse scale. On 2 November, Tutor Intelligence, an MIT CSAIL spinout, raised a $75M Series B led by Andreessen Horowitz, taking it to $117M total. Its robots work in production with remote operators the company calls tutors stepping in when they get stuck, every correction becomes training data, and customers pay per hour or per pick rather than buying anything. Neither Tutor nor Tau publishes an intervention rate, which is the number that would say how much of the work is autonomous.

AgiBot's convenience store and Unitree's chess game

From 2 to 3 November, AgiBot livestreamed six of its G2 humanoids running a convenience store in Shanghai's Xuhui district under the title "Robots Never Clock Out". They restocked shelves, made coffee and hot food, greeted customers and cleared rubbish, from the lunch rush through the night to breakfast prep. Coverage of the stream counts 1,730 tasks completed, a figure that traces back to AgiBot rather than to any audited log.

The 24 hours describes the broadcast. AgiBot has not said whether the robots stopped to charge, whether anyone helped them off camera, or what rule decided a task was complete, and reports disagree even on how many of the six struggled. Nine days earlier AgiBot had proposed a unit for robot productivity called the Worken, defined as active robots multiplied by autonomous working hours multiplied by output efficiency against a human reference. Then it ran a 24-hour demonstration and reported a raw task count.

Unitree made a larger claim with less behind it. At the China International Import Expo on 5 November, CEO Wang Xingxing said its humanoids can paint and play chess with no task-specific training, with OpenAI's GPT-6 sequencing the steps, down to swapping brush water and mixing pigments. Unitree showed video clips on stage and ran no live demonstration. Per secondary reports, the clips are accelerated and trimmed, the published chess setup pairs a board-reading model with Fairy-Stockfish, a specialist engine that picks the moves, and the painting robot is a fixed-base G1-D with two seven-axis arms rather than a walking humanoid. Nobody outside Unitree has published success rates, attempt counts, or how much code sits between the model and the actuators.

Boston Dynamics rebuilds Stretch

Boston Dynamics has had Stretch unloading boxes in paying warehouses for years, and Stretch Gen 2, described on 6 November, is a two-and-a-half-year redesign aimed at reliability instead of new capability. About 81% of the internal cabling is gone, roughly half a mile of wire, replaced in part by automotive-grade latching connectors sourced through Hyundai's supply chain. The joints rotate continuously past 360 degrees, so trajectory planning has no hard stops to avoid, and a full swerve-drive swap takes about an hour, down from roughly a full day.

The gripper change is the one worth copying. Every suction cup now has its own sensed and actuated valve, so suction runs only where contact is good. On a shared vacuum line, one cup sitting on a box edge or a torn seam bleeds pressure for the whole array, which is the case where a grip looks fine in the plan and fails in the trailer.

The fleet has passed 100,000 operating hours and tens of millions of boxes. Boston Dynamics says its most recent software release raised the whole fleet's unload rate by 24% with no hardware change, a company-reported figure, and the company notes that throughput varies with case weight and stack quality.

How much of a robot's score is the harness

A 28-author group including Pieter Abbeel released uni-robot-agent, an open trainable harness, and ran identical model weights through four harnesses on RoboDojo, a suite of 42 simulated two-armed tasks. Qwen3.8-Flash-Next scored 1.90 out of 100 in one and 31.39 in another. GPT-6 Astra went from 28.97 to 51.90.

Harness is the software around the model: it formats the camera images and the instruction into something the model can read, gives it tools such as a pointer or a depth query, and turns whatever the model writes back into joint commands. Every "model X scores Y on benchmark Z" measures a model plus somebody's harness.

The harness gets no privileged inputs, no simulator depth, no object CAD, no task-verifier feedback, only RGB, camera calibration, proprioception and the instruction. The ablations say where the lift comes from. Dropping depth costs Qwen four and a half points, and dropping the camera parameters as well collapses it to 9.13, while dropping the entire skill library costs about one point. Knowing where the camera is beats having a library of pre-written skills, which fits what goes wrong on real hardware: the model's plan is usually sensible and its geometry is not.

The paper's own headline is weaker than its table. The abstract sells Qwen's 31.39 as beating Astra's published 28.97, but Astra in the same harness scores 51.90, about twenty points above Qwen, so the comparison is against a number produced by a worse harness instead of a matched trial. The paper also leaves an inconsistency unreconciled. Section 5.1 gives Astra 55.83 with depth and 55.04 without, while Table 1 gives 56.76 and 51.90.

RoboColosseum attacks the same problem from the other end, with 26 authors including Dieter Fox, Siddhartha Srinivasa and David Hsu, plus Thomas Wolf and Pepijn Kooijmans from the Hugging Face and LeRobot side. Robots run locally in whatever room their owner has, against centrally served policies and standard protocols, giving 5,699 blinded pairwise comparisons at 124 sites across five VLA policies and three frontier agents on four bodies: the SO-101 arm, a Franka, a bimanual YAM and the Unitree G1 humanoid. Rankings barely agree across bodies and tracks, at a Kendall W of 0.24, where 1 is perfect agreement and 0 is none, and they predict post-fine-tuning performance poorly, which matters because fine-tuning on your own data is the only thing a buyer does. After a 100-demonstration fine-tune, GR00T N1.7 leads on the SO-101 while pi0.5 leads on the Franka and the YAM. On the G1 humanoid, no policy completed any of its five tasks. The abstract names neither the three frontier agents nor any per-model score, so nobody outside can check who lost.

Ai2's ToyBlocks finds the same effect in the wording, where rewriting the skill descriptions handed to an agent moved success from 24% with the defaults to 36% human-written and 63% agent-written. Taken together, every published robot leaderboard number, including a 53.9-day doubling time for frontier models driving a real arm that was fitted on one arm through one harness, has been ranking a model together with somebody's scaffolding.

A coding agent does the robot work

Two papers from a group including Daquan Zhou hand the whole job to a coding agent. In PyRUA-REAL, by Qiang Wang, Jianxin Bi and colleagues, the agent attempts a task on a real robot, keeps what happened as persistent memory, and writes reusable skills as executable code that it refines on later attempts. When it is done, the resulting program runs on its own with no agent making decisions online. It averaged 91.1% success across eight real-world tasks and 84.4% across the 42 RoboDojo tasks. That 84.4% cannot be set beside the low-20s scores frontier models post on those same tasks, because the setting differs in the way that matters: an agent that iterates, keeps notes and writes its own code is a different thing from a model asked once.

PyRUA-DataFarm, by Kewei Zhang, Juncheng Ma and colleagues, replaces the person who teleoperates a robot to collect training data. Given a data request and one calibration episode from the real robot, the agent builds the task in simulation, including the scene, an expert controller and a success judge, and a video model renders the simulated rollouts as photorealistic multi-view footage with the actions already labelled. On five real-robot tasks, policies trained on that data succeeded 93 times in 100 trials against 84 for policies trained on the same number of teleoperated episodes. Then the authors changed the room, with an unseen tablecloth and coloured lighting. The agent-trained policies held at 87. The human-data policies fell to 46.

I would have bet the other way, since synthetic data is usually the brittle option. The likely reason it held is that the renderer varies appearance from episode to episode, so the policy never has one room to overfit to, while every teleoperated episode came from the room where the rig was set up. Cost points the same way, at $0.80 per additional episode, which the authors put at roughly $107 an hour against quotes of $117 for teleoperation in China and over $290 in North America. Those are quotes the authors obtained rather than published rates, the per-episode figure only holds because the renderer is amortised, and nobody outside the group has reproduced the generalisation gap.

China blocks SoftBank's purchase of ABB Robotics

On 5 November, China's Office of the Working Mechanism for Foreign Investment Security Review, under the NDRC, prohibited the transfer of ABB's robotics operations in China to SoftBank, part of a roughly $5.375B deal for one of the world's largest industrial robot-arm businesses. The Financial Times broke it and reports the concern was that Chinese-made assets and robotics data would move to Japan and be used to train AI models, including OpenAI's, one of SoftBank's largest investments. ABB confirmed the decision and says the two sides are "working constructively on the transaction". SoftBank shares fell about 5%.

China is where the leverage sits. ABB Robotics' China orders grew 32% in the third quarter of 2026, and roughly 90% of the machines it builds for Chinese customers come out of a single Shanghai factory. The notice does not by itself end the global agreement, but the deal cannot close without those operations.

Robot trade restrictions until now have been about hardware: who may import an arm, who may sell a quadruped to a government. The asset being protected here is the recorded motion of robots doing work, and no regulator anywhere has defined what counts as exportable robot training data. The OpenAI detail is not incidental. Sam Altman said on the Sources podcast in early September that OpenAI "will definitely do humanoid robots", with a demo targeted for 2027, though no design, price or hardware partner has been shown.

Agility Robotics lists on the Nasdaq

Agility Robotics began trading as AGLT on 5 November after its merger with Churchill Capital XI closed, raising about $465M gross at a roughly $2.5B valuation. It is the first pure humanoid company whose financials anyone outside can read.

TechCrunch reported $174,000 of revenue in the first six months of 2026, mostly from related parties, against a $362M loss. Agility reports Digit deployed with nine customers including Schaeffler, GXO and Mercado Libre, and more than 65,000 operating hours. The forward book is the part to watch, with more than $300M of planned multi-year orders for Digit 5 from 30 customers Agility will not name, contingent on throughput, uptime and placement-accuracy targets it will not disclose, for a robot it has not shown in public. Prospective customers pay $500,000 for an on-site proof of concept.

VFlash drafts the whole plan at once

When you ask a vision-language model where a robot should move, it writes the answer as a visual primitive, a point, a box or a trajectory, one coordinate token at a time, so a richer plan costs proportionally more time. That makes planning one of the slowest things on a robot's own computer. VFlash, from 20 authors at NVIDIA, Johns Hopkins, HK PolyU, Purdue and HKU including Linxi Fan and Jan Kautz, cut pipeline latency on a Jetson Thor from 3,552.7 ms to 703.3 ms.

The method is speculative decoding, where a small fast model guesses ahead and the big model checks the guess, reshaped around the plan instead of a fixed token window. A tool call declares how long the output will be, the primitive's type fixes the coordinate layout, and a five-layer non-causal drafter predicts the entire primitive in one parallel pass. A structurally valid draft is accepted without checking it token by token against the target.

The surprise is that the draft is better than what it drafts for. VFlash reports 42.6% lower mean 3D displacement error on robot trajectories than its own autoregressive target, and a world action model fed VFlash plans reached 70.2% average success across four MolmoSpaces suites against 62.6% for the baseline. My read is that predicting a whole trajectory in one pass lets the drafter see the shape of the path, while the autoregressive model commits to its third coordinate before it knows where the path ends. The abstract calls its headline 25x speedup preliminary, so the Thor latency is the number worth quoting.

Say the Same, Act Differently

Some driving and manipulation models write their reasoning out in plain language next to each action, so a person or another program can read it and catch a mistake. A NeurIPS 2026 paper, Say the Same, Act Differently, by Zihao Feng, Qingzhao Zhang, Z. Morley Mao, Ramesh Govindan and colleagues, shows the text and the action can come apart. On 154 obstacle-annotated driving scenes run on NVIDIA's Alpamayo, a guided image perturbation caused a new collision while the written rationale stayed word for word identical in 52.6% of them.

The diagnostic underneath holds without any attacker. The authors measure how much the input directions that move the action overlap the directions that move the text, and get 1.87% on Alpamayo and 5.28% on InstructVLA, a manipulation model. The two outputs read almost disjoint parts of the image.

The guided attacks need gradient access, which an adversary on a public road would not have, so the overlap figure is the part that matters for anyone building a monitor: an explanation that does not change cannot tell you the action was safe. Having shipped perception onto vehicles, this matches where monitoring has to sit, on the trajectory and the actuator commands, with the model's account of itself treated as a debugging aid rather than evidence. The numbers are the authors' own and nobody has reproduced them yet.

Also this week

• Tesla will wind down Cybertruck production at Gigafactory Texas "around the middle of next year" to make room for Cybercab, Elon Musk said on 6 November. He gave no Cybercab production target, and Cox Automotive put 2025 US Cybertruck sales at 20,237, down 48% from 2024.
• Waymo is ending its Uber exclusivity in Austin and Atlanta and launching its own app there in January 2028, with co-CEO Tekedra Mawakana calling the Uber booking experience a "crapshoot" for riders. Bloomberg puts Waymo at about 4,000 vehicles and more than 500,000 trips a week across 14 US cities.
• Shanghai has written a remote-supervision ratio into law. Under rules in force from 30 October, one remote safety officer normally watches one driverless car, and an operator must get staff to a stalled vehicle within 15 minutes. No US jurisdiction publishes a ratio.

What we're watching

• Does AgiBot publish the task log for the convenience-store shift, with every charge and every human intervention, or a Worken figure for it?
• Do the robot leaderboards start reporting the harness next to the model, now that a harness change is worth more than the gap between an open model and a frontier one?
• Does any lab shipping a reasoning VLA publish a monitor that reads the actions rather than the model's written rationale?