BMW Demos New AEON 34 DoF Humanoid Robot Automation (AI NEWS)

And in terms of manipulation systems, it features bionic hands, which are equipped with modular gripping tools and scanning systems. And it can carry 15 kilograms of weight short-term, or about 33 pounds, but in terms of constant carrying capacity, it reaches up to 8 kilograms of load, or 17.6 pounds. And to power all of it, it runs on two auto-swapping batteries, each lasting up to four hours, so that when one dies, the other can take over with zero downtime, continuous operation, and this powers its 24 sensors, which include 12 cameras, peripheral, front-facing rear, as well as AR heat detection, which is all powered by this robot’s onboard AI software.

But where this robot really stands out is in its AI, so it houses three specialized onboard computers. One is a general engine, and two are dedicated AI processors for edge autonomy and mission reasoning. And it’s all running on NVIDIA’s Jetson Oren, but grained primarily in simulation using NVIDIA’s Isaac Sim and Isaac Lab. And it also learns from human demonstrations via NVIDIA’s Isaac Groot Foundation model, as well as MIMIC tools for synthetic motion data. So this is what BMW calls physical AI, and their first test deployment was in December 2025 in Leipzig, Germany.

But they’re moving on to a second deployment on the factory floor this April, 2026, with a full pilot phase happening this summer in 2026, with a focus on the robot doing high-voltage battery assembly, as well as component manufacturing. And when it comes to this robot’s cost, it’s not for sure yet, but there are rumors that they’re shooting for around a $20,000 price range. But how much would you pay for this robot? Meanwhile, robot movements are getting extreme thanks to OmniXtreme, which is breaking the generality barrier in high dynamic humanoid control.

Basically, OmniXtreme’s goal is one unified control policy that handles the full spectrum of dynamic movements. And here are the three stages while I show you this robot doing its moves. So the first stage is pre-training, and this builds the foundation. So they train a unified base policy using something called dagger based flow matching. So in simple terms, they take multiple motion tracking experts and aggregate all their knowledge into one base model. So you can think of it like learning to walk, run and balance from different coaches, and then combining it all into one athlete.

Then there’s stage two, which is post-training. This is like a reality check. And it’s interesting because first they freeze the base policy entirely and train a separate residual policy on top of it. Then that residual policy is optimized under heavy domain randomization, meaning that they throw every possible real world variation added in simulation. But the key innovation here is its power safety regularization, which is the system that actively penalizes excessive negative joint power. And it matters because during fast dynamic motions like this, joints can absorb dangerous amounts of energy that could damage the robot or even hurt somebody standing nearby.

Furthermore, they also model realistic motor characteristics like torque speed limits and actuator nonlinearities, so that simulation actually matches what the hardware can do. And this minimizes the sim to real gap problem, which brings us to stage three deployment, which is all on board and real time. So the entire inference pipeline runs on board the robot in real time means no cloud, no latency. And this enables robust as well as agile control in physical environments without depending on external compute. And this mainly matters because of safety where this approach specifically engineers against the robot hurting itself or others during its movements.

And while Omni extreme is solving how robots move their whole body, there’s an equally important problem at the other end of the scale, which is how do robot hands actually grip things precisely, because even the best vision language action models and grasping pipeline still fumble at the last step, the robot sees the object plans the grass reaches for it, and then the final grip is slightly off. And even though a few millimeters of error doesn’t sound like much in long horizon tasks where you’re chaining multiple actions together, these small arrows compound fast.

And this is what researchers are calling the last mile problem and dexterous manipulation. So now enter attack refine net, which is a framework that throws out vision entirely at the refinement stage and relies purely on touch using multiple fingertip tactile sensors, the system iteratively adjusts the robots grip based on what it feels, not what it sees, and then aligns the object to the exact desired pose. In fact, the architecture is a multi branch policy network that fuses tactile inputs from multiple fingers, along with proprioception, which is the robot sense of where its own joints are in space, and it uses this to predict precise control updates in real time.

Then in terms of training, large scale simulated data from a physics based tactile model in was used and then fine tuned with a small amount of real world data, and that sim to real combo significantly outperformed simulation only training. So what makes this really impressive is the range of what it can handle. Specifically, they tested across 16 distinct pose pairs covering pitch, roll and positional adjustments, and the system reliably nailed them all in long horizon tracking experiments where the objects pose was continuously perturbed throughout the sequence, tack refine net kept adjusting and maintaining the target grasp.

And when they tested on completely unseen objects with different shapes and thicknesses, the policy still generalize well, particularly for roll adjustments on flat objects. And to the researchers knowledge, this is the first method to achieve arbitrary in hand pose refinement using multi finger tactile sensing alone. So no vision, no external cameras, just touch. Anyways, here’s some more video showing robots doing breakdancing moves. Anyways, make sure to like and subscribe for more of the latest in AI news and check out this video here. [tr:trw].