Boston Dynamics ATLAS Demos 2026 Humanoid Robot Upgrade (AI NEWS)

And you’ll notice as this robot gains more spatial awareness, its confidence begins to increase its gait. It negotiates even at a running pace on these slick surfaces. And here we can see it sticks its first landing. Watch it roll with its wrists and pop this double flip. But there’s something that you probably didn’t notice on the first time around watching this footage. So let’s start it over again and see if you’ll catch on to what went wrong. So here’s the first time the robot is doing this double flip. Take note of its toes.

Watch them start to come apart here. And you can’t quite see the same effect in the toes in this clip. So this was probably beforehand. But after we start to see this thing come to pieces, if we watch its feet during these flips, they haven’t come apart yet. But what you’ll notice is that these toes are actually coming apart. Look at them flipping there. And then actually flying off, you can see them flying off on the ground there. So the question is, how durable is this robot really? And how many tries did they have before it began to break apart into pieces? And this leads us to the next question here, which is how durable is Boston Dynamics’ next Atlas robot that was previewed at CES 2026? So far, the only specs we have are that its battery life is around four hours and it weighs about 50 kilograms or 110 pounds with a 30 kilogram lifting capacity or 66 pounds.

And it features 56 degrees of freedom with a 2.3 meter reach or 7.5 feet and 360 degrees of motion in its joints at a height of 1.9 meters or 6.2 feet. And it weighs 90 kilograms or just under 200 pounds with built-in water resistance and an operating temperature between negative 20 to 40 degrees Celsius. But how far can the Atlas go in terms of learning, particularly with objects? Because now we’re looking at generalizable human interaction skills from human videos. This is from human X. And what they’re doing here is they’re enabling humanoids to perform these interactive tasks using human generated data.

And it starts with a basketball. And look at the way that this robot is able to maneuver almost at the level of a human. It wouldn’t be surprising if robots actually began to beat humans soon in a match of basketball. But this is only the beginning as the human X system actually integrates two co-design components, the X gen for data generation and X mimic for a unified invitation learning framework. And it allows the robot to generalize interaction skills. And it was evaluated across basketball, football, badminton, cargo pickup, and even reactive fighting, where human X was successfully able to acquire 10 different skills and transfer them to zero shot using unitry’s G1 humanoid.

And the robot even learned to do complex movements like fake outs, where it turns around and fades away jump shots, as well as carrying on interactive tasks that it was able to sustain over 10 consecutive cycles. And it learned them from just a single video demonstration from a human. And in real world experiments, human X was successful with generalization eight times more than prior methods, which is where another team push the envelope even further, as Husky just taught a humanoid to skateboard. And it’s not a gimmick, because researchers from China’s telecom tele AI Institute have developed this physics aware control system they’re calling Husky.

And it enables the unitry G one to push steer and transition smoothly on a real skateboard in both indoor and outdoor environments. So why is this hard? Well, skateboarding combines two things that robots historically struggled with. First, dynamic balance and second are non holonomic constraints. So first, when you tilt a skateboard, the trucks steer. And when you push off the ground, you shift your center of mass, and the arms and legs have to work together constantly while standing on an under actuated wheeled platform that wants to roll away. And Husky solves this problem with a phase wise learning strategy.

So the system breaks down skateboarding into three phases, pushing, steering and transitioning between them. And then it trains each phase separately, before stitching them all together with a trajectory guided mechanism for smooth handoffs. And the pushing phase uses adversarial motion priors trained on human skateboarding data to generate natural looking propulsion. And the steering phase uses physics guided tilt references based on the mathematical relationship between the board till angle and truck steering angle. So this transition phase ensures the robot can both mount the board and adjust its stance as well as switch between pushing and steering without falling.

And all the training happens in simulation across parallel environments, and then transfers directly to the real world G1 robot. And in real world testing, the robot successfully navigated outdoor terrain, made turns and maintain balance. So that’s the latest in embodied AI. But there’s also been a breakthrough in generative AI as anthropic just released its Claude Opus 4.6. And it’s their most capable model yet. So now Opus 4.6 plans more carefully sustains agentic tasks for longer. Plus, it operates more reliably in large code bases and catches its own mistakes through better code review and debugging.

And it gets a 1 million token context window in beta, which is a first for anthropics Opus class models. And the benchmarks are backing it up because Opus 4.6 achieved the highest score in terminal bench 2.0 for agentic coding. And it leads all frontier models on humanities last exam, which is a very complex multidisciplinary reasoning test. And in long context performance, it saw a massive jump with Opus 4.6 scoring 76% compared against its sonnet 4.5 at 18.5%. And on top of all this anthropics API features adaptive thinking where models can decide when to use deeper reasoning with four effort levels for developers to balance intelligence versus speed as well as compacting context.

And don’t forget to like and subscribe for more of the latest in AI and robotics news. [tr:trw].