Simplexity Robotics Demonstrates Robot-Run CNC at IROS 2026; BASAL Intelligence Raises Funding for Action-Native Embodied AI

At the IROS 2026 Tech Talk on Sept. 29, 2026, Feng Zongbao, reinforcement learning lead at Simplexity Robotics, described the company's target of using robots to manufacture robots. The hardest part, according to the talk summarized by Leiphone, is precision CNC operation: workpieces must be picked, loaded into chucks, inserted, removed and placed with a clearance of only 0.5 mm. Reflective metal surfaces, weak textures and similar-looking parts make visual alignment difficult, and critical contact interfaces are not always visible.

Simplexity Robotics said its full-stack embodied AI combines self-developed multimodal models, a closed-loop data flywheel, a modular hardware platform and the SimpleClaw developer platform. Its hardware lines include the i7 Pro wheeled robot, the iX wheeled humanoid, modular force-controlled robotic arms with the Dexter dexterous hand or gripper, and data-collection gloves for scaling human demonstrations.

The company's technical approach uses three complementary methods. SimpleWAM generates action chunks from multimodal observations; DRAM, or Delta-Rule Recurrent Associative Memory, handles long-horizon context with a fixed-size memory matrix; and DPE uses a separately trained evaluator to score candidate actions while keeping the policy frozen. According to the talk, the combination addresses action generation, memory and action selection in the CNC system.

In nominal trials, Leiphone reported, the fused method succeeded 15 times, compared with 0 times for Query-concay and 8 times for a π0.5 SFT baseline. On precision CNC insertion, the combined method achieved a 100% task success rate. The model was trained with only 600 real robot trajectories. Simplexity Robotics demonstrated a complete workflow in which one robot autonomously tends two CNC machines, including picking workpieces, loading the chuck precisely, and removing and placing finished parts.

In the second development, Leiphone's Whale Rhino reported exclusively that Li Jianxiong, the first PhD graduate from Tsinghua AIR, has founded BASAL Intelligence. The company has received investment led by 5Y Capital, with Ivy Capital, Linear Capital and WestSummit Capital participating. Its core team includes chief scientist Zhan Xianyuan, an associate professor at Tsinghua AIR, and seven young researchers who have long collaborated with the team.

According to BASAL, before forming the founding team, core members had received high-paying offers from ByteDance Top Seed, Nvidia and Xiaomi, as well as overseas PhD admissions. They collectively gave up annual packages totaling nearly RMB 40 million, with Li personally giving up a maximum annual package of nearly RMB 10 million.

BASAL said it will not follow the mainstream pretrained foundation model plus post-training deployment paradigm. Instead, it is betting on an action-native in-context embodied foundation model, aiming to let robots learn and adapt on site without retraining when scenes change. The company argues that action information from physical interaction is difficult to describe fully and accurately through language, and that current language-centric VLA routes have inherent limits in fine manipulation and cross-scene generalization.

The company's goal is for robots to face new scenes, tasks or embodiments unseen in training and, with a small number of human demonstrations and autonomous attempts, complete rapid learning and adaptation within minutes. It is developing two core technologies: X-VLA, a cross-embodiment model that uses less than 1,000 hours of pretraining data and, according to BASAL, achieved leading results in third-party evaluations and won the overall championship at the IROS 2025 AgiBot Robot Challenge from more than 400 teams; and ODEWorld, an efficient latent world model that models continuous physical time to support long-horizon physical representation and in-context learning.

Leiphone's Whale Rhino reported that BASAL is among the earliest Chinese teams to systematically explore cross-embodiment rapid adaptation and in-context embodied learning. The report also noted that pursuing an action-native route rather than a language-centric one raises short-term communication costs with customers, who are more familiar with natural-language robot interfaces. The company's next challenge is to validate on-site learning within minutes across more real scenes, embodiments and task types.