Tesla's Optimus can now sort objects on its own

Remember when Tesla first announced its humanoid robot with a person dancing in a spandex suit? Those days are officially over.

Over the weekend, Tesla shared a new video of its Optimus robot on X, showing the humanoid operating entirely on its own to sort physical objects using an end-to-end trained neural network.

Video in, controls out

In the demonstration, Optimus starts by self-calibrating its limbs using joint position encoders alongside its onboard cameras. According to Tesla, this allows the robot to accurately understand where its body is in physical space.

Once calibrated, Optimus begins sorting blue and green wooden blocks into matching trays with impressive dexterity.

The real highlight comes when a Tesla engineer steps in to mess with the robot. The worker physically moves the blocks around mid-task, but Optimus doesn't freeze or fail. Instead, it dynamically tracks the changes and keeps sorting the blocks into the right trays.

In another sequence, Optimus accidentally places a block on its side. Without any human prompt, it immediately notices the mistake, reaches back, and flips the block upright before moving on.

Tesla capped off the video with Optimus holding a one-legged yoga pose, pairing the update with a link to its robotics careers page to recruit motion planning, controls, and computer vision engineers.

The actuator bottleneck

This update marks clear progress from the tethered prototype seen at Tesla's 2022 AI Day and the hand mobility clips shared in May.

However, getting Optimus off the demo stage and onto factory floors remains an uphill battle. During Tesla's Q2 earnings call, CEO Elon Musk admitted that Optimus production had fallen well behind internal targets.

The main culprit was a severe shortage of off-the-shelf actuators suitable for humanoid robotics. Tesla was forced to design and manufacture its own custom actuators from scratch, aiming for eventual volume production.

Why it matters

We think the most significant takeaway here is architectural: Tesla is applying the exact same end-to-end visual neural network approach from its Full Self-Driving efforts directly to robotics.

While traditional robotics often relies on rigidly hardcoded routines for perception, trajectory planning, and joint actuation, Optimus processes raw video input to generate motor commands directly. The proof is in its dynamic adjustments: when an object gets moved or lands awkwardly, the robot course-corrects rather than crashing out.

For the robotics industry, Tesla's willingness to build custom hardware from the ground up could drastically lower actuator manufacturing costs over time once volume kicks in. But it's also clear that stacking wooden blocks on an empty table is still a controlled showcase. The true test will be surviving the chaotic, unstructured environment of an actual manufacturing line.