Boston Dynamics New Atlas GR3 AI Robot Hand: 5 Engineering Upgrades

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Boston Dynamics New Atlas GR3 AI Robot Hand: 5 Engineering Upgrades · Channel: AI News · 08:01 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI models, Robotics and autonomous systems

Summary

Boston Dynamics’ upgraded Atlas GR3 hand has 13 degrees of freedom and is designed around practical skills such as pinching, gripping tools, and manipulating objects for manufacturing work. The video explains how proprioception, backdrivability, and accurate simulation can make the hand both more dexterous and more durable, and describes design trade-offs such as omitting a pinky. It then shifts to AI agent and model announcements, covering Manus 2.0 and its tools, OpenAI’s DOTS and GPT 6.1 Sol, and Google’s Gemini 4 Argon. These announcements are presented as part of growing competition in AI agents and professional workflows.

Key points

The Atlas GR3 hand increases the previous GR2’s seven degrees of freedom to 13 and is slightly larger and stronger than an average human hand.
GR3 is designed for behaviors such as precise pinching, tripod grasping, and holding a tool while independently pressing its trigger.
The hand’s target tasks relate to work Atlas may perform in Hyundai manufacturing facilities, including tool use and assembly.
Proprioception and backdrivability let Atlas sense forces and motion at its fingers, while allowing joints to move out of the way on impact.
Atlas simulation models kinematics, dynamics, and contact dynamics; the team worked to align simulation and hardware from both directions.
The team chose not to include a pinky because its added dexterity was not worth the additional degrees of freedom, size, and power draw.
Manus 2.0 introduces the Cascade agent harness, Cloud Computer, triggered automations, Manus Studio, and the Q personal-agent app.
OpenAI’s DOTS are described as agents with individual cloud computers and browsers, app connections, preference learning, and controls for sensitive actions.
The video says GPT 6.1 Sol is a lower-cost model for coding and computer use, while Google’s Gemini 4 Argon targets long, complex professional workflows.

Chapters

Start	Chapter	Summary
00:00	Atlas GR3 hand: design and dexterity	The video introduces the GR3’s increased degrees of freedom and its intended grasping and tool-use skills.
01:35	Proprioception, durability, and simulation	It explains backdrivability, the role of accurate simulation, and the design decision to leave out a pinky.
03:46	Manus 2.0 and its agent workspace	The video outlines Manus’s new agent harness, persistent computing, automations, Studio tools, and Q app.
05:56	OpenAI agents and models	It describes OpenAI’s DOTS agents and GPT 6.1 Sol’s reported capabilities and pricing.
07:27	Google Gemini 4 Argon	The video presents Argon’s workflow focus, reported benchmark results, and rollout plans.

Quotes

this new GR3 hand now jumps to 13 total degrees of freedom. — 00:00

So back drivability makes this possible — 02:09

their biggest decisions still came down to trial and error — 03:14

you set it up with a single prompt. — 04:19

its output limit jumps from 64,000 to 1 million tokens — 07:27

Concepts and entities

Boston Dynamics, Atlas, GR3, proprioception, Cascade, Manus Studio, Q, DOTS, GPT-6 Astra, GPT-6.1 Sol, Gemini 4 Argon

Full transcript (original language)

00:00 Boston Dynamics just unveiled its upgraded Atlas hands built on five key engineering choices that let it pull off several new dextrous feats. And the team calls this new generation the natural companion to the newest Atlas designed with mass manufacturing in mind by favoring simplicity wherever possible. except for the degrees of freedom, as the previous GR2 hand featured just seven degrees of freedom, while this new GR3 hand now jumps to 13 total degrees

00:30 of freedom. And in terms of its size, the GR3 is slightly larger than the average human hand while also being a bit stronger. But joints only matter if the hand can use them, which is where dexterity comes in. And the team built around specific behaviors from day one, like a pinch grasp that’s precise enough to roll something small against the thumb, or a tripod grasp that holds on to an object with more strength while still being able to move it, or else

01:01 gripping a tool’s handle while independently pressing its trigger, too. And it’s this kind of range that matters because the hand stays close to being fully anthropomorphic while maintaining a small gap between a human demo and the hand’s own performance, which makes it far easier to simulate. And importantly, Boston Dynamics chose these tasks because they represent skills that Atlas actually needs in order to be deployed in Hyundai’s manufacturing facilities. And those same reorientation tasks are also central to how the GR3 hand learns.

01:35 And once mastered, the foundation for both tool use and assembly is effectively built, which are two things that Atlas needs for real work. Furthermore, reinforcement learning trains the robot’s entire body, and the team is actually extending that to teach the arms and hands to manipulate objects directly. But this only works if the hand is simulatable and transparent to force and motion. And that transparency comes down to propriception, which is the biological term for sensing where your body and muscles are in space,

02:09 whether they’re moving and how much force is applied. And the same is true for Atlas, where any force the environment puts on its fingers translates back into something that Atlas can sense, and any motion it applies goes out just as cleanly. So back drivability makes this possible and it does it double duty by first enabling dextrous movement and second saving the hardware by letting a joint accelerate out of the way on impact instead of absorbing it which is exactly what would otherwise wear down and break the hand

02:41 over time. So Atlas’s upgraded propriception leads to not just a more capable hand, but also a more rugged one. But none of this works unless the physical hand and its digital twin match. So, simulation for Atlas covers three layers. First, kinematics defines geometry where the links and joints sit. Second, dynamics cover friction, torque, and backlash. And third, contact dynamics handles how the hand physically interacts with the world. In fact, Boston Dynamics said getting sim to real to match meant working from both

03:14 directions at once instead of just tuning simulation to match the hardware. But despite all of the team’s engineering, their biggest decisions still came down to trial and error, as there were no clean answers, just competing objectives weighed through simulation, 3D printed mock-ups, and testing designs themselves. And this ultimately led to the team deciding the hand didn’t need a pinky because that extra dexterity wasn’t worth the three additional degrees of freedom or the size or the extra power draw. And while Atlas is getting hands, AI agents are getting something just as big, their own

03:46 computers. As Manus just launched Manus 2.0, and it starts under the hood with the new in-house agent harness called Cascade. It keeps projects lightweight at the start and only adds specialized capabilities when a task actually needs them. And the payoff here is 23.2% fewer tokens, 28.2% faster completion, and 32% lower cost compared to the previous system. So, it does the same work faster and cheaper. Then there’s Cloud Computer, which is a dedicated environment for projects that need to keep running. That means always on

04:19 servers for multiplayer games and 24/7 automations. And speaking of automations, they now go beyond scheduled tasks to even triggered workflows. Emails, ad changes, calendar events, Slack messages, and notion updates can all set Manus in motion, and you set it up with a single prompt. But the biggest addition is Manus Studio, which is a desktop app built as a professional workspace for documents, spreadsheets, slides, websites, code, games, and video. And it debuts two new environments. The video editor creates a

04:51 first cut in minutes with an editable timeline where clips, images, text, motion graphics, and audio are all separate components that you can swap and reprompt. And then it also targets short ads, AI UGC tutorials and vlogs. And an alchemy mode combines video and code generation for more creative direction. Then gamedev gives you playable templates from the start, plus an edit panel where you can change code, assets, art, and scenes, even editing one asset without touching what surrounds it. And multiplayer runs

05:22 through cloud computer, so you skip the hosting and networking headaches. Furthermore, Manus Studio also adds remote control and computer use, so you can reach your desktop from your phone and have Manus carry out tasks in an authorized workspace with access to your files and apps. So, it’s like your desktop can work while you’re away. And finally, there’s Q, a brand new standalone app for personal agents where each agent gets its own identity, a dedicated email, phone number, wallet, and computer. And the agents can join group chats and then hand tasks off to each other. And Q can even work with QR

05:56 code restaurant systems to order food or manage a queue. So currently it’s in early access free with the invite code meet Q on web, desktop, and mobile with iOS coming soon. But Manis isn’t the only one giving agents their own computer because OpenAI just introduced DOTS. And dots are AI agents powered by GPT6 Astra that work on your behalf around the clock, each with its own cloud computer and browser. They connect over 4,000 apps through plugins, learn

06:26 your preferences over time, and can even start tasks before you ask using readonly access. And an autoreview system checks sensitive actions against your instructions. And you decide what’s allowed, what needs approval, and what’s blocked with an activity view showing the background work. Plus, DOTS lives in chatGPT, Slack, and Teams with SMS coming soon. And while dots run on GPT-6 Astra, OpenAI also just launched GPT 6.1 Sol, which is a cheaper model built for

06:57 coding, computer use, and professional work. And on coding benchmarks, it matches Astra at roughly 1/5 if of the cost. And on OS World 2.0, it lands within 2.1 points of Astra at 17th the cost. and factual errors at low reasoning effort dropped to 32% with price being $2 per million input tokens and $10 per million output tokens. But these releases from OpenAI is heating up the competition as Google just released Gemini 4 Argon which is built for long

07:27 complex professional workflows and its output limit jumps from 64,000 to 1 million tokens with Argon setting a new high on DeepSWE version 1.1 at 77.9% and leading automation bench at 51.3%. It also handles 800,000 line codebase migrations and ties for first on CWE bench. So pricing starts at $2 per million input tokens and $10 per million output tokens. And it’s rolling out first to trusted cyber defenders, then

08:01 to paid API customers and Google AI Ultra subscribers.