R1 Humanoid Robot Beats AI Experts With This ($4900)

Then there’s the R1, which has 26 degrees of freedom, but also no Dextra’s hands. Then finally, there’s the R1 EDU, which has up to 40 degrees of freedom, because it features 5-fingered Dextra’s hands. And looking at this robot jumping up and down on one fist, tell me down below in the comments whether or not you can make out four fingers balled up with a thumb. Because that would lead us to believe that we’re actually watching the R1 EDU. And in terms of the hardware of all three of these models, they all have a max arm torque of about 2 kilograms or 4.4 pounds.

But when it comes to vision, the R1 Air features a monocular camera, whereas both the R1 and the R1 EDU feature monocular cameras. And when it comes to power, all three feature a smart lithium battery with quick release and a battery life of about one hour. But they differ when it comes to their AI, because the R1 and R1 Air both feature eight core CPUs with 10 tera operations per second. But then the R1 EDU features an additional dedicated Jetson Oran GPU. And all three models stand a similar four feet or 123 centimeters tall, but they slightly differ in weight with the R1 Air weighing about 59 pounds or 27 kilograms.

And the R1 EDU weighing about 29 kilograms or around 64 pounds. And looking at the website, it seems like they come in two separate colors. But it’s hard to tell what real world tasks these robots can perform aside from these type of entertainment moves. But comment down below which model is most worthy of the price. Also, tell us besides acrobatics, what could these unitry robots do in the real world? And tell us whether you’d rather lease or purchase these robots. But unitry has another robot that’s performing real housework autonomously in collaboration with BitRobot and HuggingFace with the launch of HIW 500, which is a massive open source data set featuring over 500 hours of humanoids operating in the wild.

And moving past structured labs, this initiative captures real world household chaos across different home scenarios throughout Southeast Asia. And it was done using unitry’s G1 robot to track bi-manual interaction and mobile manipulation across 161 unique sub-task labels, including sweeping floors, cleaning, and stocking refrigerators. And the data set actually logs multimodal inputs at 30 frames per second, combines them with 480p RGB and IR stereo streams using its head camera and wrist-mounted sensors alongside its 29 degrees of freedom joint state metrics, its IMU telemetry, and fine-grained language annotations. And it’s all hosted on HuggingFace with its supporting raw robot-as-a-system bag and MCAP recordings, as well as LaRobot formats to accelerate global end-to-end imitation learning, with an expectation to expand their library later this quarter.

And while data sets like HIW 500 are teaching robots what to do in households, another critical question remains, how do they move fluidly without falling? Well, Tsinghua University and Galbot just bridged this gap by introducing humanoid GPT, which is a transformer-based framework that achieves zero-shot motion tracking. And the team shattered previous limits by training the model on an unprecedented 2 billion frame motion corpus, and it utilizes a new harmonic motion embedding metric, which balances data diversity, and then it distills multiple reinforcement learning experts into a single generalist transformer. And it’s also deployed on Unitree’s G1, with humanoid GPT mastering dynamic, unseen real-world tasks, including soccer, digging, and dances, with rock-solid balance.

And it’s optimized via TensorRT with 1.5 millisecond inference speed that proves scaling laws can apply directly to humanoid agility. Next, Paxini and GeelyAuto just teamed up using the Terra-1 humanoid to help in its factories. Here you can see it’s using its binocular depth cameras with 60 pose estimation, with its ultra-dextrous hands to be able to pick up this plug, and it has an accuracy of just 0.01 Newtons and takes over 1 million samples per second in order to insert this wiring harness, which is an ultra-dextrous task. Next, PNDbotics just demoed the world’s first humanoid robot successfully climbing a 1 meter tall box.

And this was a test of full-body multi-joint coordination and core strength using the Atom humanoid. And you can see here that it was able to quite easily negotiate its way up this box and stand. Meanwhile, in generative AI, there is a brand new open source model that is competing directly with Anthropic’s Fable 5. And this is from a company called Z.AI from China, and it’s an open source model called GLM 5.2, which is a massive 753 billion-parameter open-weighted model that’s designed specifically to master long-horizon autonomous software engineering. And it was released under an unrestricted MIT license on Hoggingface, with GLM 5.2 addressing the major scaling pain of long-context token costs by deploying a custom architecture called IndexShare.

And by sharing a single indexer across every four sparse attention layers, this model slashes per-token flops by 2.9 times, making its 1 million-token context window computationally viable for hosting on local codebases. In fact, in benchmarks, it scored nearly neck-and-neck with Fable 5. Plus, GLM 5.2 also includes specialized dual reasoning modes, high and max, to prevent plateaus during multi-step tasks. And it’s available right now via API if you don’t have physical GPUs to run it locally on your computer, with open router charging just $0.95 per million input tokens and $3 per 1 million output tokens.

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