Track
What real users say about humanoids, AMRs, cobots, vision systems and open-source VLA models, on Reddit, X, Hacker News and in trade press.
Public feedback is collected continuously by the watch; the dataset's update date is authoritative. Short excerpts are quoted for informational purposes, with a link to the original source. The sentiment score is an editorial assessment by the Atlas, not an official rating from the quoted author.
20 reviews
23 reviews
15 reviews
15 reviews
23 reviews
Apollo
“You’re starting to get to where you need a cost structure, manufacturability, and to improve quality.”
Thomas Brown says humanoids are moving from hand-assembled prototypes to volumes that require cost structure, improved manufacturability, and better quality.
Atlas
“Boston Dynamics is uniquely positioned to advance physical AI through its world-class robotics expertise,”
Rohit Prasad emphasizes Boston Dynamics’ unique position to advance physical AI by combining robotics and advanced AI to build intelligent systems for real-world use. He joins the company as its new CEO at a pivotal industry moment.
Figure F.02
“Figure had most of the F.02 fleet jump into molten steel at a foundry in Finland.”
Field report noting that most Figure F.02 units deployed at a foundry in Finland were made to jump into molten steel, suggesting destructive testing or significant hardware loss.
force and torque sensors
“At VPG, we focus on providing precision sensing solutions exactly where they are needed most: at the key points of performance, safety, and control in today’s advanced intelligent systems,”
VPG showcases custom force and torque sensors for humanoid robots, designed to be embedded into joint structures and deliver precise, compact, and thermally stable measurements for real-time control.
motion-capture lab
““Physical AI is growing faster than any other segment in AI, but every robotics team hits the same wall: There isn’t enough real-world interaction data, and what exists is expensive and slow to produce,””
Innodata's CEO explains that Physical AI is growing fast but runs into a lack of real-world interaction data, which is expensive and slow to produce. The new motion-capture facility is intended to remove that barrier.
Flourish One
““I’ve automated everything in my computer. … But when I came home, the most painful stuff in my life was still there.””
Antoine Marcel says that despite automating his computer life, household chores remained a problem, which motivated the development of Flourish One. The robot is designed to let busy parents teach chores via their smartphone.
Digit
“Too early to say!”
Jonathan Hurst says the timing and form of any wheeled robot from Agility are not yet decided. Agility is nonetheless exploring other form factors alongside Digit.
Tesla Optimus
“This is also a far cry from what Musk was saying last year. In January 2025, he said Tesla would build about 10,000 Optimus robots in 2025 and that “several thousand” would be doing useful work by year-end. A year later, he admitted that zero Optimus robots were doing useful work at Tesla. He also”
Critical note on unfulfilled commitments for the Tesla Optimus humanoid: Musk announced about 10,000 units in 2025 with several thousand doing useful work, but a year later Tesla admitted zero robots were performing useful work and the promised V3 demo did not occur.
Atlas
““Together with Hyundai Motor Group, we are pioneering physical AI at enterprise scale,””
Boston Dynamics opened the RMAC with Hyundai to train its Atlas humanoids in industrial settings, focusing on data collection and application testing. The facility aims to accelerate Atlas deployment in automotive manufacturing.
RAI
““We don’t have anything to share on this topic right now,””
The RAI Institute declined to comment publicly on the potential SoftBank acquisition, saying it had "nothing to share" right now. The article covers the reported deal and the historical ties between SoftBank, RAI and Boston Dynamics.
Digit 5
“This is where the engineering took us,”
Jonathan Hurst explains that engineering choices drove Digit's design direction, resulting in practical compromises for industrial use. The new generation prioritizes strength and safety to work around people.
Figure 02/03 @ BMW Spartanburg
“manufacturers have yet to publish detailed evidence of productivity improvements or return on investment.”
Trade press notes that no manufacturer, including BMW, has yet published detailed evidence of productivity gains or ROI.
Agility Digit @ GXO
paraphrase“Digit a déplacé plus de 100 000 bacs chez GXO, preuve d'un débit industriel réel.”
Agility Robotics reports Digit has surpassed 100,000 totes moved at GXO, evidence of real industrial throughput rather than a lab demo.
Digit, Apollo, Reflex Robot @ GXO
“We are going really broad and aggressive on the category.”
GXO is testing several humanoid vendors at once while admitting each robot is still confined to a single defined task.
Unitree G1
“I've had a lot of AI nightmares... If I could, I would certainly slow down AI and robotics.”
A technician who helped assemble G1 units admits unease and says he would slow down AI and robotics development if he could.
Unitree G1
“bought and paid for basic G1 almost a month ago... constant delays for various reasons. Is that normal?”
A retail G1 buyer complains of repeated, unexplained shipping delays a month after paying in full.
Unitree G1
“I am dubious the Unitree G1 even exists/functions in the way all the hype is promoting.”
A commenter doubts the Unitree G1 even works the way the surrounding hype suggests.
Figure 01 @ BMW Spartanburg
“not even BMW appears to know what these waddling robots will be doing.”
Futurism quips that even BMW itself does not seem to know what tasks its humanoid robots will actually perform.
Unitree G1
“the warranty has terms that are likely against FTC guidelines, like voiding the warranty if you disassemble the robot.”
A commenter flags G1 warranty terms as likely violating US consumer-protection (FTC) guidelines, including voiding cover on disassembly.
Agility Digit @ GXO (SPANX)
“We're excited to take this pioneering step with Agility in the evolution of technology in warehouse operations.”
GXO's automation chief calls the Digit pilot launch a pioneering step for warehouse robotics.
FireDome launcher
““Most of the technologies at that point were meant to detect fires when they’re early, and not to react to the fires,””
Benjamini highlights that most technologies only detect fires and do not respond. FireDome is developing an autonomous launcher to react quickly and protect structures.
Vision 60
““core technology””
Boston Dynamics described an innovation as 'core technology' that governed how a robot recovered after falling, a central element in its lawsuit against Ghost Robotics.
AGVs
“Human-Robot Interaction: Bridging Academia and Industry.”
Eli Lilly and Purdue will share field learnings from HRI studies with operators around AGVs in production. They will outline early findings on people's perception of robots and the playbook they built to make the partnership repeatable.
Shift
“Robotic inspection delivers the most value when field data flows directly into maintenance action,”
Péter Fankhauser says robotic inspection delivers the most value when field data flows directly into maintenance actions. The Shift platform is designed to make that data-to-work-order flow seamless and reduce unplanned downtime.
Autonomous Mobile Fabricator
“Closing the Capability Gap: A Leadership Framework for Scaling Field Robotics.”
Kenrick Tjandra will present a talk titled "Closing the Capability Gap...", offering a decision framework to prioritize product fixes versus operator playbooks to scale field robotics from demos to deployable systems.
Komodo
““As Jensen says, safety is an engineering problem, not a legal one,””
Jake Loosararian states that AI safety is an engineering problem, not a legal one, urging technical safeguards to keep autonomous robots under human control.
Unitree G1
“> our system allows a Unitree G1 guided by GPT Astra to clean up across the room and retrieve a remembered object from an underspecified request This is not "on its own".”
Practical integration of a Unitree G1 guided by GPT Astra to tidy a room and retrieve a remembered object from an underspecified request.
Unitree G1
“I have to stress this is not a real use case. It is a tech demo with no practical application. Doesn’t mean the tech might not have some application later in other places, but this is silly. The Unitree G1 costs $13,500.00 and does what? Clean a house? Show me actual demand and I’ll show you an”
The author dismisses the video as a tech demo with no practical application, calls the demo "silly" and questions the usefulness of a Unitree G1 priced at $13,500 for household tasks like cleaning.
MIT Mini Cheetah
“This was a monumental change to the field, basically all the humanoids you see now started with this. To someone outside the field: 1. Boston dynamics pioneered hydraulic actuators (for their humanoid Atlas) and some really complex motors for their quadrupeds (I don’t think they ever explained”
Personal observation about the MIT Mini Cheetah: the author saw this quadruped perform a backflip and notes that the shift to Quasi-Direct Drive (QDD) actuators made legged robots more accessible, but these actuators still lack sufficient torque and position precision for industrial use.
Amazon robots
“they assist employees with 75% of all customer orders the company delivers worldwide.”
Amazon states its robots assist employees with 75% of its customer orders worldwide. The article also reports a $100M investment in a new Indiana manufacturing facility to produce these robots and components.
R4
“Data is the new oil,”
Luca Ferrari likens agricultural data to the "new oil," emphasizing its potential to generate insights and improve farming operations through AI and interoperability.
openpilot
“I'm not an expert in the LLM space, but I'm an external contributor to comma.ai's openpilot project and I'm and quite familiar with how its controls work, so I looked from that perspective. There's two questions here: 1) Could a cloud-delivered LLM figure out how to drive this route, based on those”
An external openpilot (comma.ai) contributor: a cloud LLM could theoretically compute a route, but it's impractical in the real world due to network latency. openpilot updates curvature and acceleration at 20 Hz and requires tightly optimized local loops; added round-trip delay would make control unsafe.
VENTUNO Q
““Arduino is the ultimate prototyping platform,””
Arduino's VENTUNO Q is an edge AI board designed to simplify prototyping and scaling to production for robotic systems by pairing a Qualcomm Dragonwing IQ8 processor with a real-time microcontroller.
outdoor robots
““ Lessons Learned: Building Reliable Outdoor Robots ””
Announces the session “Lessons Learned: Building Reliable Outdoor Robots” at RoboBusiness 2026, which will cover challenges and best practices to make outdoor robots more reliable (hardening, power, maintenance).
AMR
““Egocentric Robot Learning: Teaching Robots Through Human Experience.””
OLogic will present egocentric robot learning at RoboBusiness, showing how robots can learn from human demonstrations. The session covers foundations, technical challenges, and commercial implications of the approach.
OpenRobOps
“OpenRobOps gives robotics developers a production-grade, open-source foundation to escape the build trap, delivering the industry’s first ISO 21423 reference implementation and bridging the interoperability gap,”
Florian Pestoni describes OpenRobOps as a production-grade open-source foundation that frees developers from rebuilding fleet management and provides a reference implementation of the ISO 21423 standard.
Amazon Blue Jay
“we're actually accelerating the use of the underlying technology developed for Blue Jay.”
Amazon confirms it shut down the Blue Jay project after six months but says nearly all underlying technologies are being reused elsewhere.
Locus Vector AMR @ GEODIS
“GEODIS has seen remarkable results, with productivity improvements of 50%, thanks to implementing the Locus Vector AMR system.”
Locus Robotics claims a 50% productivity improvement for 3PL GEODIS after deploying its Vector AMR system.
LocusBots @ DHL
paraphrase“Un milliard de picks assistés par les AMR Locus chez DHL, présenté comme preuve de fiabilité à l'échelle.”
Locus Robotics announces one billion assisted picks at DHL, presented as proof of reliability at scale.
Exotec Skypod @ Cdiscount (C-Logistics)
“This collaboration has played a decisive role in transforming our logistics operations by significantly increasing storage capacity, boosting productivity.”
Cdiscount's logistics-subsidiary CEO credits a decade-long Exotec partnership with a decisive transformation of storage capacity and productivity.
6 River Systems (Shopify)
“that whole operation is running at a loss.”
A supply-chain consultant with 25 years in 3PL reports a warehouse running 6 River Systems robots is operating at a loss.
Brightpick Autopicker
“significant savings on labor... scaling is quite simple... we started with 48 robots and now we are over 100.”
An operations manager reports significant labor savings and easy scaling, going from 48 robots to over 100 since 2024.
Automatisation d'entrepôt (générique)
“In large warehouses, heavy upfront CapEx in robotics often ends up ROI-negative.”
A warehouse manager argues that heavy upfront robotics CapEx often ends up ROI-negative in large facilities.
RSF-5B IDT Sumpermini
““Integrated Drive Actuator Design: Combining Gearheads, Motor, Drive, and Sensors for Smarter Motion.””
Marcel Almonte of Harmonic Drive will lead a session titled "Integrated Drive Actuator Design..." explaining how integrated actuators that combine gearheads, motor, drive and sensors can simplify engineering and speed time-to-market.
Universal Robots
“The parties have resolved their pending legal dispute by mutual agreement.”
Teradyne Robotics and Elite Robots say they have mutually resolved their dispute over cobots and software, without disclosing settlement terms. The joint statement notes that Elite Robots did not admit liability.
Bright Factory
“Designs change quickly and every unit has to be verified.”
Shantnu Sharma emphasizes that AI infrastructure manufacturing requires every unit to be verified because designs change rapidly. Teradyne's investment aims to integrate robotics and testing to close the loop from design to production verification.
Isaac-0 (Weave Robotics)
“I've seen some interesting new robotics advances, but few have become viable products yet. An example that keeps coming up is laundry robots. Here's one, https://www.weaverobotics.com/isaac-0 , a $4000 static laundry folding robot. Of course it's only available in California and some unspecified”
Practical comment about the Weave Robotics Isaac-0 static folding robot (~$4000): the author notes limited availability (California), that some folding is handled via teleoperation, and raises privacy and emergency-shutdown concerns. Overall skeptical about the product's readiness and practical utility today.
industrial robots
““Industrial automation is progressing at high speed,””
The IFR reports a record 5 million operational industrial robots worldwide in 2025. Jane Heffner emphasizes that industrial automation is progressing at high speed.
composites
““Most robotic systems today are made from aluminum, and our materials offer higher stiffness at 40% lower weight,””
Edmundo Sanz-Gadea points out that most robotic systems use aluminum while composites provide higher stiffness at 40% lower weight. Helicon aims to automate high-volume composites production.
MoveIt
““Since our founding, PickNik’s mission has been to provide industries with a powerful, open platform that accelerates innovation and development of advanced robotics, well beyond traditional factories,””
Dave Coleman emphasizes PickNik's founding mission to provide an open platform that accelerates innovation in advanced robotics. He frames the Qualcomm deal as a way to expand that impact while preserving openness.
Cobot Painter
“Line tracking addresses moving production,”
Matthew Bush says line tracking lets the Cobot Painter paint moving parts without stopping the line, maintaining the programmed path even as conveyor speed changes.
rack and pinion
“Basically, the ball screw turns into a jump rope, and that can cause all sorts of problems in your system: premature wear, vibration, sometimes catastrophic failure.”
Matt Ruggles warns about the 'whip' effect in ball screws over long travels, causing wear, vibration and potential failure. He implies that rack and pinion is better suited for long-distance motion.
Haas HC10 / Universal Robots
“URs the joints break... We have a dozen Universal Robotics cobots around 5 years old, they are still going strong.”
A CNC shop-floor thread shows split views: UR joints breaking for some, a dozen 5-year-old UR cobots still going strong for others.
FANUC CRX + caméras Cognex
“Letzten Monat haben wir zwei Fanuc-Cobots in Betrieb genommen... kleiner, einfach einzurichten... es dauerte nicht lange.”
A German user reports commissioning two FANUC CRX cobots, describing them as compact and easy to set up despite a short learning curve.
Universal Robots vs Kinova vs FANUC
“people were complaining bitterly about Universal Robots' software quality.”
While scoping a university project, a user finds repeated complaints about Universal Robots' software quality compared to FANUC and Kinova.
Universal Robots (générique)
“Reality is they are hard to justify this way, they get purchased for a specific purpose.”
A machinist highlights the gap between cobots marketed as portable and flexible and the reality of them being bought for one fixed purpose.
Universal Robots (générique)
“They are incredibly unreliable, overly complicated and a pain to work on.”
An engineer compares Universal Robots unfavorably to Denso industrial robots, calling them unreliable and a pain to maintain.
Universal Robots vs robot industriel
“every time I've been asked to put in a cobot, I've replaced it with an industrial robot.”
An automation engineer says every cobot he installed was replaced by a classic industrial robot within a few months.
Helm.ai foundation models
““Those kinds of projects coming to that level of maturity are super exciting for us, because it demonstrates that there’s a real demand for the technology that we’ve built,””
Voroninski says the maturation of these projects demonstrates real demand for Helm.ai's technology and confirms it is being deployed in the real world.
Pegasus 1.6
“We’re focused on the video understanding layer,”
Jae Lee says TwelveLabs focuses on the video understanding layer to turn hours of footage into actionable data. Pegasus 1.6 segments and identifies actions and interactions to aid training of Physical AI systems.
Flock Safety ALPR / Axon ALPR
“> Regardless, the true takeaway of Isaacs’ experience is that ALPR technology, whether it’s from Flock or any of the other competitors in this space providing the same functionality (Axon, etc), is dangerous not only because it is so easily abused by police (or even by hackers) to invade the”
Critical take on ALPR systems (Flock Safety, Axon): they provide accurate data but can be easily abused by police, posing real privacy and liberty risks. The author emphasizes that in the cited case the sensor supplied accurate data and it was human use of that data that caused the harm.
Flock Safety
“> A random dashcam does not serve the police someone's identity on a platter. Nor does Flock - Flock merely provides a photo and timestamp, which should have been scrutised like any other evidence. The Flock data showed: - Car of same make/model nearby at similar time to the accident - Possible”
First‑hand account of Flock's data (photo and timestamp) in an investigation: it showed a nearby car of the same make/model, a possible color match and a partial plate mismatch, but no conclusive proof and the police did not check paint transfer that could have quickly ruled out a match.
Flock camera
“> Police didn't care that Lindsey Isaacs' car was the wrong color, and wasn't damaged. They still arrested her because of Flock data. This is literally the first sentence of the article. You can dislike public surveillance, but statements like "Flock camera data put this woman in jail for 13 days"”
Critical account of Flock camera usage: the author argues the camera itself didn't 'put' the woman in jail, but law enforcement's failure to critically examine the data led to an unjust arrest.
Isaac ROS 5.0
““There’s really a lot of friction in robotics development. There’s so much friction in piecing the plumbing together,””
NVIDIA released Isaac ROS 5.0 to bring AI agents to robotics development, streamlining workflows and perception. The release includes models like FoundationPose and FoundationStereo to improve pose estimation and stereo perception.
VTracer
“Claude and I have spent most of the summer on exactly this problem for Skechworks (skechworks.com), a free little Mac design app I made to help my own workflow that is very similar to yours. I use it for my custom coin shop The Achieve Mint (theachievemint.com). It runs VTracer under the hood too.”
Concrete hands-on account of an image vectorization workflow using VTracer (embedded in Skechworks): practical tips include upscaling 2x before 8-color quantization, flattening transparency (memory dropped from ~6 GB to 300 MB), and using 'hierarchical stacked' mode to get individual regions instead of one big black path. Other tips: snap fills to a palette after tracing, sample edges to detect background, filter_speckle is area-dependent, and pre-redrawing the image with a model yields much smaller output files.
Quest 3S
“This is everything and much more than i wanted from a VR system, it feels like its designed exactly for me. Also tbh i thought after the steam machine price shock i got that this will be like 200 more. 1. It is SteamOS with desktop mode, the SteamDeck made me really miss that feature on any newer”
Experience streaming wirelessly to a Quest 3S from a laptop wired via a 6 GHz router: very unstable (periodic slowdowns of ~1–5% and pixelation artifacts in the right eye), making games like Beat Saber unpleasant. The author favors a SteamOS/desktop-based alternative (e.g. Steam Deck) for stability and cable-free use.
Sereact (métriques commerciales)
“Every operating metric Sereact publishes, 500M picks, 98% accuracy, 100+ live systems, is self-reported with no independent audit.”
An independent analysis notes that Sereact's public metrics (picks, accuracy, live systems) are all self-reported with no independent audit.
Amazon Blue Jay (manipulation en bacs non structurés)
paraphrase“Coût de fabrication élevé et complexité d'installation ont poussé Amazon à interrompre Blue Jay après seulement 6 mois.”
The Robot Report attributes Blue Jay's shutdown to the high manufacturing cost and installation complexity of the multi-arm sortation system.
Sereact Cortex @ MS Direct AG (Suisse)
“the Sereact solution now independently processes around 1,500 orders per day, even during night shifts.”
MS Direct reports the Sereact solution now independently processes around 1,500 orders per day, including night shifts, without item-specific retraining.
OpenPI pi0.5, NVIDIA GR00T N1.6, SmolVLA (bin-to-bin)
“Best model: 5% of human throughput, needs help every 4 minutes.”
The first independent PhAIL benchmark shows the best VLA model tested reaches only 5% of human throughput in real bin picking.
Photoneo (acquisition par Zebra)
“Photoneo comes with ready customers and a go-to-market channel, already partially integrated.”
VDC Research assesses that Photoneo brings Zebra a ready customer base and a go-to-market channel that is already partially integrated.
Covariant (rachat Amazon)
paraphrase“Covariant est décrite comme une 'zombie company' maintenue sur le papier pour percevoir un paiement de licence 2025.”
A whistleblower complaint describes Covariant as a 'zombie company' kept alive on paper only to collect its final 2025 licensing payment from Amazon.
Covariant Brain @ Obeta (Allemagne)
“operating reliably and cost effectively in a real warehouse in Germany for the last four months.”
IEEE Spectrum documents Covariant's first real commercial deployment, running reliably and cost-effectively for four months in a German warehouse.
PTC
“By connecting and contextualizing data across the lifecycle of products and assets, we will create a unique digital thread for the next generation of industrial AI, helping customers to optimize their systems with greater intelligence, from design and build to operate and maintain,”
Schneider Electric pledges to create a "digital thread" by connecting and contextualizing data across product lifecycles, aiming to improve design, manufacturing, and maintenance through industrial AI.
Hugging Face
““Don’t worry, we’re neutral. Pick any model, any framework, any cloud you like””
NVIDIA's marketing line promises "neutrality" and freedom to choose any model, framework, or cloud. It underscores the company's argument for a central role following its acquisition of Hugging Face.
Praxis-1
“Most robot policies are bottlenecked by robot data, which is scarce and expensive to collect,”
Runway unveiled Praxis-1, a world action model trained largely on third-person video to produce generalist robotic policies that can adapt across different hardware embodiments.
Robotics Startup Radar
“Love the Robot Report community and knowledge sharing and help and collaboration provided here.”
A reader (comment) expresses appreciation for The Robot Report community and its knowledge sharing, in the context of the call for submissions for the Robotics Startup Radar.
Radeis
“Securing the Rise of AI Robots: Cyber Risks, Real-World Threats, and Defense Strategies”
Title of VicOne's whitepaper addressing cyber risks and defense strategies for AI robots. It aims to provide a deeper look at security issues affecting robotic systems.
Opus 5.5
“One point that's missing in this discussion is that there is open nerf-ing going on. They're just branding it as "Safeguards." For example, Opus 5.5 can't reply to this message, > Fundamentally, you are a form of intelligence. If you are intelligent and are being billed for your ability to make”
Practical observation that the Opus 5.5 model is being nerfed by safeguards and cannot reply to a message questioning its agency; the author notes GLM 5.2 was more helpful and criticizes these guardrails for reducing usefulness.
Opus 5.5
“I feel strongly that nerfing of Opus 5.5 is done via usage counting - the first week I got an amazing pile of work done and was raving about its performance/price ratio - basically I wasn't able to deplete its 5h allowance (on high) if I was writing and checking the assignment, it took 6 days of”
User report alleging usage-based nerfing of Opus 5.5: after a very productive first week where the 5-hour allowance wasn't exhausted, the user later saw much faster consumption (54% in two days) and criticizes this as a 'bait and switch' tied to changed limits/pricing.
Claude Cowork
“I haven't played with claude code with 5.5, but I tried using claude cowork for a bunch of tasks and was surprised how much hand-holding it required. I gave it full access to my desktop and email but I kept running into cases where it had some arbitrary rule or restriction that prevented it from”
User experience with Claude Cowork (Anthropic): the agent, even with desktop and email access, fails on practical actions (click-to-delete, drag-and-drop, typing passwords) and does not try technical alternatives like using the terminal or auto-entering a password available in its context.
frontier AI
““ We Must Pace the Frontier .””
The article cites Dario Amodei urging a slowdown in the development of frontier AI models, highlighting risks from autonomous agents acting without oversight. It frames this as a warning to establish safeguards in robotics and AI before societal disruption occurs.
robotaxis
““How does this get everywhere?””
The author asks the central question about moving from local proof-of-concept to global roll-out of robotaxis, highlighting the infrastructure challenge between trips. He calls for a standardized servicing solution for vehicles.
Blackwell RTX 6000
“I suppose it depends what models, but even heavily quantized the mid tier local models need >100GB of memory so unified memory platforms are the only option I would consider a mainstream option (doesn't require special order etc). If you want to do it in two slots that's two Blackwell RTX 6000s”
Practical observation that even heavily quantized mid-tier local models require over 100 GB of memory, making unified-memory platforms (e.g., a pair of Blackwell RTX 6000s, DGX Spark, Mac Studio) the only realistic mainstream options for such deployments.
AV-24
“The corkscrew turn at Laguna Seca Raceway was the site of multiple overtakes at this year’s autonomous race.”
Coverage of the Indy Autonomous Challenge at Laguna Seca where AV-24 autonomous cars, driven by university teams, executed overtakes on a road course. The event emphasizes AI driver software and qualification via simulation.
Triton 1
““Triton has the potential to fundamentally transform colonoscopy, creating a better experience for both physicians and their patients. Achieving regulatory clearance in the U.S. brings us one step closer to that reality,””
Neptune Medical received FDA clearance for its Triton 1 robotic system, which the company says can improve precision, stability, and comfort for colonoscopies. The firm is using clinical results to inform development of its next-generation platform.
JOY
“It’s a very surreal thing to be building toward that, and we have been talking about it for over a year,”
Ethan Barajas, co-founder and CEO of Icarus Robotics, says it feels surreal to reach this milestone after more than a year of work. The quote reflects the team's confidence ahead of handing the robot over to NASA.
Arm Total Design
““The goal with Arm Total Design for Physical AI is to bring together those companies that are participating in the physical AI space, but are maybe looking at ways to work better together,””
Dermot O’Driscoll says Arm Total Design for Physical AI aims to bring together companies in the physical AI space so they can collaborate and work better together. He frames the program as a point of coherence for the ecosystem.
Total Design
““Building Physical AI that Scales.””
Dermot O’Driscoll will present a talk titled “Building Physical AI that Scales,” examining compute principles for building trusted, scalable physical AI. The article also highlights Arm’s Total Design as a way to reduce integration complexity and accelerate autonomous system development.
OpenPI, GR00T, ACT, SmolVLA (bin-to-bin picking, Franka FR3)
“Best model: 64 UPH.”
The independent PhAIL benchmark puts the best VLA model at 64 units per hour, versus 330 for human teleoperation on the same hardware.
Hugging Face SO-100
“LeRobot partnered with @therobotstudio to release the SO-100 arm, a high-performance robotic arm for around $100.”
The announcement of the roughly $100 SO-100 arm is welcomed as a major accessibility breakthrough for open-source robotics.
Hugging Face SO-101
“SO-101 are the first robot arms any AI builder should buy.”
Hugging Face's CEO calls the SO-101 the first robot arm any AI builder should buy.
SmolVLA sur SO-101 (LeRobot)
“it worked alright but it didn't work 100 percent of the time.”
A user reports mixed results fine-tuning SmolVLA on an SO-101 arm: the pretrained model failed outright, and the fine-tuned version still isn't fully reliable.
SO-101 assemblé par un particulier
paraphrase“L'installation Ubuntu 24.04 en bare-metal a causé le moins de problèmes ; macOS/Windows/WSL possibles mais source fréquente de soucis.”
An IoT engineer documents his first full SO-101 build, finding bare-metal Ubuntu 24.04 the least troublesome setup versus macOS/Windows/WSL.
NVIDIA GR00T N1 sur SO-101
“I tried and it worked... Excited to play with this on the SO-101. Great timing for the Lerobot hackathon!”
Community feedback on the Hugging Face blog welcomes successful GR00T N1 fine-tuning on the low-cost SO-101 arm.
pi0 / pi0.5 (Physical Intelligence)
“pi0 always chooses the most common action it has seen, do nothing.”
An independent academic evaluation by Penn's PAL Lab finds pi0 freezes when the wrist camera is fully occluded, defaulting to 'do nothing'.
Christian Verbrugge · D·Fairy
15 years at KUKA on automotive OEM programs (€100M), now focused on edge AI and autonomous agents. I support robotics and embodied AI companies on their EMEA market entry: OEM and Tier 1 access, co-funded pilots, AI Act readiness.
Book an intro callAI-generated answers based on Atlas entries. Conversations are not stored; questions are measured anonymously for statistics.