TLDR Hardware 2026-09-09

Robocurve, a public benefit company backed by YC, has opened a community bounty program to build reproducible open-source evaluation benchmarks for physical AI and autonomous robot systems. To make these benchmarks happen, it is giving away $20,000 and a pair of I2RT ABC Box robot arms to 25 teams who propose the best research benchmarks. Robocurve hopes to accelerate the development of Physical AI by increasing the number and diversity of dexterous robotics benchmarks.

Chinese regulators are tightening approval for humanoid robotics startups seeking to go public, following a volatile trading debut from sector leader Unitree Robotics, as they try to cool a sector described as sizzling but rife with copycats and little real technological breakthrough. The China Securities Regulatory Commission's tighter posture is expected to affect other humanoid IPO hopefuls, reportedly including names such as AGIBOT and Galbot.

Xiaomi's 18 Fold launched in China a full 24 hours ahead of Apple's rumored iPhone Fold, packing a wide-style, passport-sized chassis with a 5.38-inch outer OLED and 7.58-inch inner display, both hitting 4,000 nits peak brightness with a 1-120Hz LTPO refresh rate and, notably, an RGB sub-pixel structure that's a first for any foldable. Xiaomi prioritized battery over thinness, fitting a 6,000mAh cell (with 67W wired charging) into a 5.02mm-unfolded body that's still slightly thicker than Samsung's Galaxy Z Fold8.

This open-source repository presents a comparative benchmark evaluating four popular LiDAR odometry and SLAM algorithms—KISS-ICP, DLIO, FAST-LIO2, and GLIM—on live aerial drone flight data collected with a Livox Mid-360 LiDAR and Pixhawk flight controller. Running directly on a Jetson Orin NX embedded computing platform, the project compares trajectory accuracy and map reconstruction quality against millimeter-grade VICON ground-truth data. These empirical findings provide developers and perception engineers with actionable performance data to choose and tune state estimation algorithms for compute-constrained autonomous platforms.

Data collection for Physical AI is an expensive problem blocking the development of more advanced models. To get around this bottleneck, researchers at Rice University have developed RoboTok, an automated data engine that indexes, filters, and processes large-scale human demonstration videos from the internet to train manipulation policies. By extracting hand trajectories and object interactions directly from uncurated web video, the system learns complex, multi-stage manipulation skills without requiring physical teleoperation. This approach showcases how massive online video datasets can be converted into scalable training data to overcome physical data collection bottlenecks.

Orion is an open-source, 12-degree-of-freedom quadruped robot designed, fabricated, and programmed from scratch using ROS 2, custom 3D-printed components, and serial bus servos. The repository provides complete mechanical CAD models, kinematic calculations for dynamic walking gaits, and step-by-step firmware integration guides. This project provides engineering teams and builders with an accessible reference architecture for prototyping low-cost legged locomotion platforms.

NASA is sending a seven-axis robot into low-Earth orbit, and they are looking for researchers who would like to test their code in the challenging environment of microgravity. The NASA Fly Foundational Robots (FFR) Mission is a competition where teams develop autonomous software to show off the potential benefits and challenges of space robotics. Do you think you have what it takes? The last day to submit an application is September 23.

They say any sufficiently advanced technology is indistinguishable from magic. Sean Tsai, an engineer at Foxconn, has built an interactive robot that performs classic sleight-of-hand magic tricks for human audiences. It integrates computer vision to track spectator choices and low-latency motor control to execute precise, synchronized mechanical deceptions in real time. It illustrates how combining high-speed visual tracking with fine motion planning can significantly enhance interactive entertainment and human-machine collaboration.

IonQ published a 70-page resource estimate concluding that a trapped-ion quantum computer with 19,397 physical qubits, built on its “Walking Cat” architecture, could compute a 256-bit elliptic-curve discrete logarithm on secp256k1, the curve securing Bitcoin and Ethereum, in about 25.7 days per attempt.

Agility Robotics, one of the early leaders in humanoid robotics, reported $1.8 million in revenue alongside a net loss of $114.7 million for 2025 as it prepares to go public via a SPAC merger.

Quantinuum secured $100 million in CHIPS and Science Act funding, the only trapped-ion architecture selected in this funding round, to tackle a core scaling problem: the specialized ion traps, control electronics, lasers, and optics behind its quantum processors get harder to manufacture consistently as systems grow larger.

Huawei has become the de facto coordinator of China's push to build domestic lithography machines.