New Tesla Optimus Killer? $314 Robot Hand Does This (AI NEWS)

So expect the hand joint count to potentially jump from 15 to 20 or more degrees of freedom, approaching human-level manipulation capabilities. Then we can also likely expect enhanced AI processing power. So the current Turing chip in the iron humanoid uses a 40-core processor, so it’s probably going to be upgraded to between 60 to 80 cores, or it’ll have a multi-chip architecture. Given that they’re calling 2026 the chat GPT moment of humanoids, they could have some new hardware in the pipeline. Next, in terms of improved perception systems, their previous Eagle Eye system could evolve alongside their automotive updates, because remember XPeng also makes smart cars.

And this could include higher resolution mapping, faster processing speed, and possibly additional sensor modalities. Plus XPeng could also integrate new thermal or depth sensors that they’re already using in their car fleet. Then in terms of mobility and speed, the upgraded humanoid will likely integrate new reinforcement learning policies for better walking, faster walking, better terrain balance. And finally, their newest robot could integrate a new level of autonomy for expanded operation periods, better task planning without human intervention, and more sophisticated decision making in unseen environments. But what if you could skip the full humanoid and just get the arms with the hands? This is what Mimic Robotics dual hand five-finger dexterity system is already working on, and it’s performing in the real world.

But let’s break down what makes this so interesting. So starting first with the hand itself, looking at the hardware specs, it has 21 joints and 16 degrees of freedom. So that’s quite a bit of articulation, but less than a human hand. But the tenon drive design is clever because it keeps the weight down to just one kilogram, while also handling up to seven kilograms of payload. And in terms of precision, they say that it can operate within plus or minus a half a millimeter at the fingertips. And for this, the wrist mounted cameras are key.

So notice these dual wide angle RGB cameras that are positioned at the wrist. This gives the system visual feedback right where the manipulation is happening. And as for how it actually learns, they’re training on multiple data types simultaneously, including images, language and other robot demonstrations. And more importantly, human video. So the real innovation seems to be these wearables that they mentioned, they’re using them to capture how humans perform tasks, and then they’re feeding it directly into the training pipeline. And this controls the system. It’s called high frequency generative control, which means that in practice, the AI is making predictions and adjustments constantly generating smoother movements rather than following a rigid pre-programmed path.

And the results are 93.3% success rates on tasks that the system hasn’t specifically trained on. And this is huge because it suggests actual generalization, not just memorization. And that 33.3% performance boost from self-correcting behavior, that’s the system catching and fixing its own mistakes mid task, which is genuinely useful. And they’re mounting these hands on Franca Amica panda arms, and you can get it in one or two arms stationary or mobile. And it’s using a hardware agnostic approach, which means that you’re not locked into one configuration, customers can adapt to their needs.

And currently they’re testing pick and play sorting and assembly insertion. And they’re hoping that these foundational manipulation tasks can translate across industries for more general purpose use cases. But let’s take a look at another breakthrough from Tether IA as they just unveiled their arrow hand open. And it’s a fully functional robotic hand that costs less than a smartphone at just $314. It’s completely open source. And the problem that it solves is that for years, researchers and developers face this critical barrier, which was that proprietary robotic hands cost thousands of dollars.

And this made advanced manipulation research out of reach for most labs, universities and independent innovators. But the arrow hand open changes this equation entirely. So in terms of the technical innovation, what makes this hand remarkable isn’t just the price, it’s the engineering at under 400 grams. This hand is lighter than a can of soda. So it delivers professional grade performance, but it uses a clever tenant driven architecture. So instead of requiring individual motors for each joint, the design routes low friction cables from shared motors to multiple joints simultaneously, which mimics how human tendons work in the body and the hand features 16 degrees of freedom, seven power joints working on nine passive couplers that recreate natural hand movements.

And the thumb alone has three active degrees of freedom, which enables for opposition pinching and in hand object manipulation. And with 10 newtons of force per fingertip, the hand is strong enough to pull a trigger on a drill, or it can manipulate other everyday objects. But perhaps most impressive is the hands 100% back drive ability, meaning that you can physically move the fingers by hand without resistance. And this is important because it absorbs impacts like shocks, which makes interactions safer for both objects and humans. And it supports multiple control modes, including position, torque and inline force sensing to enable truly adaptive manipulation.

And once again, it’s open source, meaning that it’s accessible to everyone. But tell me in the comments if you would use an open source robot hand like this. And in terms of full humanoid robots that are already working, Dubot Adam just released some new footage here. This robot costs $27,500. And it features 28 degrees of freedom and a neuro driven dexterity system. And this is driven with a transformer based AI architecture and dual RGB cameras. And the system features 200 hertz of high frequency control, meaning it updates 200 times per second.

Next for the robots imitation learning and reinforcement learning systems, it consumes 42% less energy as compared with the traditional bent knee robot. And for its AI brain, it’s using the ROM one, this is the robot operator model one combining general and industry specific AI models, featuring 100 million parameters with 24 hertz end to end control. And this allows for autonomous decision making in unpredictable or unstructured environments. And the atom features a total of 41 degrees of freedom with 47% more. But the atom comes in two models, the atom trainer model and the atom max variant.

The one in the video is likely this atom max with 41 degrees of freedom, which is 47% more articulation than the base model. And the atom max features these dexterous 12 degree of hands, 60 frames per second full HD vision with the Intel real sense depth sensing cameras, and VR or mixed reality teleoperation for markerless tracking. And on board, it features 1500 tera operations per second of AI computing for real time edge processing, as well as 360 degree LiDAR for environmental mapping. Last but not least, here’s the G1. And this is without the any to track this is the G1 pulling a car.

But like and subscribe, check out this video here. And thanks for watching. Bye. [tr:trw].