At ICRA 2026, the IEEE International Conference on Robotics and Automation, six leading Chinese robotics firms — lidar, tactile, simulation, humanoid, force control, and adaptive arm makers — each gave industry keynotes arguing for the same fixes to perception, contact control,…
At ICRA 2026 in Vienna, six of the most-watched Chinese robotics companies gave six different industry keynotes and landed on the same three fixes. The most revealing of those three is the one they cannot yet claim to have solved: the "last millimeter" problem, the gap between a robot's learned model of the world and the variable, compliant surface of the real thing.
ICRA, the IEEE International Conference on Robotics and Automation, is the field's largest annual research meeting. The 2026 edition ran in Vienna in early June, and the official industry-keynote program featured six Chinese-headquartered firms: lidar maker RoboSense (速腾聚创), tactile-sensing startup Pacini Perception (帕西尼感知), simulation-infrastructure vendor Lightwheel Intelligence (光轮智能), humanoid-robot builder AgiBot (智元机器人), force-control specialist Skywiser Robotics (天机智能), and adaptive-arm maker Flexiv (非夕科技). A Lei Feng Wang (雷锋网) synthesis of the six talks, published July 1, 2026, organized the talks around three intersection themes: a sensor-layer push to fuse vision, depth, and motion in one device; the "last millimeter" contact-and-force problem; and the data flywheel that links simulation to the real world.
The convergence reads less like a coordinated playbook than like a shared bottleneck. None of the six companies is trying to win on the same product. They are trying to escape the same wall.
That wall has three faces. The first is perception. RoboSense argues that post-processing pipelines accumulate error, so its next-generation lidar, the upcoming E2 series and the 孔雀 ("Peacock") chip, is designed to fuse RGB, depth, and motion at the sensor level rather than after the fact. Pacini Perception (帕西尼感知) and Skywiser (天机智能) layer tactile and contact-force sensing on top of vision, and Pacini says it is building a 500-person team and five planned tactile-data factories to feed that hardware. Flexiv (非夕科技) threads contact force upward into the control stack rather than treating force as a downstream correction.
The second face, and the one the talks circled most carefully, is what several speakers called the "last millimeter." The phrase captures a stubborn problem in manipulation. A robot can plan a grasp to within a centimeter using vision, but the last few millimeters of contact are dominated by compliance, friction, and surface variability that vision cannot see. Skywiser's TORA-ONE humanoid platform leans on multi-point force control to absorb that variability, and Flexiv's own ICRA 2026 preview, corroborated by Robotics and Automation News and TipRanks, frames compliant force control as the company's answer to the same problem. Pacini's dexterous hand, the DexH13, packages 16 degrees of freedom into a tactile sensor array so that contact itself becomes the input.
The third face is the data flywheel. Lightwheel Intelligence (光轮智能) sells simulation-to-reality infrastructure rather than robots, and is closely associated with NVIDIA's Newton physics engine, a parallel, GPU-based simulator designed to close the sim-to-real gap at training time. AgiBot (智元机器人), the most foundation-model-oriented of the six, has staked its platform on large-scale pretraining and a "deployment is training" loop in which robots in the field keep generating data that retrains the model. The AGIBOT World Challenge at ICRA 2026 is structured around that premise. Both bets assume that the gap between a simulator and a factory floor can be closed with more data and better continual learning. Neither has been demonstrated at scale.
The pattern across the six talks is the news. When six well-funded companies, each with a distinct product, all hammer the same three problems in the same week, the most honest reading is that the problems are not yet solved. The "last millimeter" gap is not a research curiosity. It is the reason a humanoid that looks right in a demo video can still fail to pick up a coffee cup on a soft, slightly tilted surface. The data flywheel is not a marketing slide. It is the only credible path from today's brittle, hand-engineered manipulation to a robot that can learn on the job.
What to watch: whether RoboSense's E2 series and Peacock chip ship with sensor-level fusion on by default, whether AgiBot's continual-learning loop survives contact with real customers, and whether Lightwheel's simulation infrastructure, and its Newton-engine bet, gets adopted by enough of the field to make cross-company training data realistic. The next ICRA, in 2027, will be the first venue where these bets can be evaluated against each other rather than presented in parallel.