Alongside the news that Boston Dynamics is going to let its robot dog, Spot, out of its laboratory for the first time, the company has released a new video of Atlas, its spectacular bipedal robot that’s previously been seen doing everything from parkour to backflips. In this latest video, Atlas does a small gymnastics routine, consisting of a number of somersaults, a short handstand, a 360-degree spinning jump, and even a balletic split leap.

What’s most impressive is seeing Atlas tie all these moves together into one pretty cohesive routine. In the video’s description, Boston Dynamics says that it’s using a “model predictive controller” to blend from one maneuver to the next. Presumably each somersault gives the robot a fair amount of forward momentum, but at no point in the video does it seem to lose its balance as a result. Amazingly, Atlas is able to roll gracefully along its back without any of its machinery getting squashed or tangled.

A new machine-learning model can now translate neural activity into readable sentences with only three percent error.

“Proprioceptive Control of an Over-Actuated Hexapod Robot in Unstructured Terrain,” by Marko Bjelonic, Navinda Kottege and Philipp Beckerle from Technische Universitat Darmstadt and CSIRO, Brisbane, Australia was presented at IROS 2016 in Daejeon, South Korea.

In recent years, researchers worldwide have been trying to develop sensors that could replicate humans’ sense of touch in robots and enhance their manipulation skills. While some of these sensors achieved remarkable results, most existing solutions have small sensitive fields or can only gather images with low-resolutions.

A team of researchers at UC Berkeley recently developed a new multi-directional tactile sensor, called OmniTact, that overcomes some of the limitations of previously developed sensors. OmniTact, presented in a paper pre-published on arXiv and set to be presented at ICRA 2020, acts as an artificial fingertip that allows robots to sense the properties of objects it is holding or manipulating.

“Our lab recognized the need for a sensor for general robotic manipulation tasks with expanded capabilities than current tactile sensors,” Frederik Ebert, one of the researchers who carried out the study, told TechXplore. ‘“Existing tactile sensors are either flat, have small sensitive fields or only provide low-resolution signals. For example, the GelSight sensor provides high resolution (i.e., 400×400 pixel) images but is large and flat, providing sensitivity on only one side, while the OptoForce sensor is curved, but only provides force readings at a single point.”