In the fast-evolving world of humanoid robotics, Tesla's Optimus program grabbed headlines with revelations about its massive computational backbone, as a new permit for Giga Texas' Cortex 2.0 datacenter—dedicated to Optimus training—confirms capacity for up to 200 MW of power, enough to sustain around 162,000 U.S. households or fuel tens of thousands of GPUs for intensive robot learning. This scale underscores the raw infrastructure demands propelling humanoid capabilities forward, with Elon Musk firing back at critics by asserting that ">Ironically, Optimus, FSD and AI will provide all the things this moocher demands"—hinting at broader societal impacts from Tesla's robotics push.
Over at Figure, CEO Brett Adcock shared rare optimism after months of grueling development, noting from a recent office visit that they've cracked a "transformative AI problem" after twice-daily standups since summer: ">We’ve been working on a transformative AI problem for nearly 6 months... humanoid robots are so hard. Both the dynamics and high dimensionality make it such a highly complex system level problem. Today was one of our best days so far." He emphasized that "all the ingredients needed for general purpose robotics are here at Figure," signaling imminent breakthroughs in full-spectrum humanoid versatility.
Meanwhile, Apptronik capped the year on a high note, earning a spot as a 2025 Top Tech Startup from SDCExec for its Apollo humanoid's strides in warehouse logistics, celebrating "development milestones and Apptronik’s commitment to helping humans in warehouse settings."
Pushing the frontiers of robotic dexterity and long-horizon tasks, Chris Paxton illuminated key technical levers from recent research. He highlighted how ">given a strong enough base of robot tasks data, it becomes much, much easier to leverage human data"—unlocking cross-embodiment transfer—while "action chunking—predicting a short trajectory segment instead of just the next robot action"—curbs compounding errors in motion planning. Paxton also spotlighted emerging solutions like MemER, which infuses memory into policies for feats like object search, addressing a "crucial" gap in today's mostly "memoryless" robot demos. Rounding out the insights, he praised scaling's magic where ">beyond a certain scale things start to work" and the reliability boost from practical 3D sensing over RGB alone, potentially fast-tracking real-world deployments.
The global humanoid race intensified with fresh looks at China's ecosystem, courtesy of Tuo Liu. Ant Group—Alipay's parent—is honing humanoid chefs, while NOETIX unveiled bionic humanoids tailored for museum receptionists and guides. Even niche demos shone, like the AgiBot humanoid's archery attempts—nailing near-perfect form from one angle despite a miss—showcasing dexterity in sports. Liu, mapping humanoids worldwide and dubbing Shenzhen the "center of drones and AI robotics," fielded bold comparisons as the "Jack Ma of AI robotics," fueling visions of an "Alibaba of robotics."
Industrial robotics deployments underscored hardware maturity, with Kawasaki Robotics powering high-speed material handling via its Kawasaki BX130X arm—handling polyethylene rolls at up to 6 per minute with smart recognition and traceability courtesy of partner Tiesse Robot SpA. Similarly, FANUC America touted a cobot dual-system for fabrication, blending welding and plasma cutting in a 96" x 48" envelope at 8x faster speeds alongside Northtech Automation's HS Fabmaster.
These threads—from Tesla's power-hungry training hubs and Figure's engineering triumphs to dexterous research hacks, Chinese innovation bursts, and rugged industrial installs—paint a robotics landscape hurtling toward ubiquity, where hardware grit meets scalable intelligence for warehouse aides, public guides, and beyond.
Top comments (0)