New Robot Hardware With 20 DoF Beat Optimus? + Claude Opus 4.7 (AI NEWS)

And as for its form, the entire unit measures just 201 by 75 by 50 millimeters, or about 8 by 3 by 2 inches, making it compact enough to integrate onto standard robotic wrist mounts without adding significant bulk or inertia to the system. But light doesn’t mean fragile, because the WUJI hand supports a maximum static grasping load of 10 kilograms, or just over 22 pounds. And in vertical grasping configuration, it holds 5 kilograms, or just over 11 pounds. On top of this, its maximum hanging load is 5 kilograms. Then each finger supports a hook grip of 3 kilograms, or about 6.6 pounds independently.

And when it comes to range of movement, the WUJI hand provides 20 active degrees of freedom distributed evenly across all 5 fingers, with 4 degrees of freedom in each of them. Each finger also contains 4 joints, all serially connected through direct drive rotary mechanisms. And this kinematic configuration eliminates the nonlinear relationships that typically compromise control accuracy in cable driven, or tenon based hand designs, allowing every joint to move predictably, and every movement to translate directly to the fingertip. On top of this, the thumb is fully opposable, and includes lateral swing support, enabling the kind of cross palm manipulation patterns that are essential for tasks requiring fine repositioning of objects within its grasp, with its maximum grasping diameter reaching a total of 100 millimeters, or just under 4 inches, which is wide enough to wrap around most handheld objects that would be found inside of a kitchen.

Then there’s precision, with every joint in the WUJI hand being powered by brushless motors operating at millisecond level response times, and the drive mode uses a self locking rotary direct drive system, meaning its joints hold their position without continuous power draw, which is a critical feature for sustained grasping tasks where energy efficiency really matters. And its control architecture runs on field oriented control vector algorithms, with dual encoder feedback, so this combination delivers a positioning repeatability of give or take 1 millimeter, or give or take 0.04 inches, ensuring that the hand returns to the same configuration for consistent accuracy across repeated movements, and its control interface operates over USB 2.0, which makes its integration straightforward for research and development platforms, but a hand is only as good as it is sensitive.

At each fingertip, force sensors register pressures as small as 15 Newtons, or about 3.37 pounds of force. This tactile feedback layer gives the system real time awareness of contact forces, allowing the control software to modulate grip strength dynamically, and whether the hand is holding an object that’s rigid or fragile, the sensor data feeds directly into the control loop, enabling the kind of adaptive force management that separates functional robotic hands from truly dexterous ones. But would you trust one of these robots inside of your kitchen? Currently, both robotic hands are integrated into a single platform through Liber AI’s control stack, with the only drawback being preparation time currently, but will this change in the near future? But dexterous hands are only one piece of the puzzle, because a hand needs a body, a body needs a brain, and a brain needs somewhere to work, and skilled AI just secured all three.

So the company just announced its acquisition of the robotics division Zebra Technologies, previously known as Fetch Robotics, to deploy its omni-bodied brain across warehouse environments in the real world. And this move pairs one of the most battle-tested warehouse robotics platforms in the industry with a foundation model designed to generalize across any robot body, quadrupeds, humanoids, tabletop arms, or mobile manipulators, all without retraining from scratch. And nearly every physical product consumer’s purchase passes through a warehouse at some point, with fulfillment being a complex, high-volume process operating inside massive facilities where orders must be picked, packed, and shipped with speed and accuracy.

And most existing warehouse robotic solutions rely on classical approaches, but significant portions of the pipeline, particularly tasks involving the movement of objects between receptacles, remain bottlenecked by human labor. And skilled AI’s thesis is that these limitations stem from the brittleness of traditional robotics deployments, because conventional systems are programmed around specific hardware which are tuned to exact kinematics, meaning that when you change the robot, the programming starts over. But the skilled AI brain was built to eliminate that dependency entirely. It operates without prior knowledge of the robot’s body form and generalizes across embodiments at the architecture level.

And here’s what Zebra brings to the table. So Zebra’s robotics division contributes more than just hardware. Its symmetry fulfillment orchestration platform already coordinates tasks between robots and frontline workers using real-time data from Zebra wearable devices. And by integrating the skilled brain into the foundation means that robot warehouses move beyond executing fixed instructions and move towards making informed adaptive decisions. And this acquisition also feeds directly into skilled AI’s data flywheel, with more deployments across more environments generating more diverse operational data, which in turn trains and improves the omnibodied brain, creating a compounding cycle of capability.

But while robots are getting smarter, large language models may be heading in the opposite direction, as many users are actually reporting a dramatic downgrade in performance and intelligence between Opus 4.6 a week ago and Anthropic’s newest release, Claude Opus 4.7. So Anthropic says the model delivers notable improvements over Opus 4.6 in advanced software engineering, with particular strength in complex, long-running tasks that previously required close human supervision. Early testers also report that Opus 4.7 handles multi-step workflows, follows instructions more precisely, and independently verifies its own outputs before reporting results. But other users are reporting that Opus 4.6 has been significantly dumbed down, with Opus 4.7 seemingly not even achieving what 4.6 did about a week ago.

As for vision capabilities, Anthropic says they’ve also received a substantial upgrade, with Opus 4.7 accepting images up to 2576 pixels on the long edge, roughly three times the resolution of prior Claude models, enabling finer detail in tasks like reading dense screenshots, extracting data from diagrams, and working with pixel-level references. Additionally, Anthropic also introduced a new X-High effort level for finer control over reasoning depth, task budgets in the public beta for managing token spend, and an ultra-review command in Claude code for dedicated code review sessions. Opus 4.7 is now available across Claude products, the API, Amazon Bedrock, Google Cloud’s Vertex AI, but tell us in the comments below if it feels like Anthropic nerfed their models to charge the same amount of money for less of a product.

Anyways, like and subscribe and check out this video here for more of the latest in AI and robotics news. [tr:trw].