Industrial robotics enters the dark factory era

Dark factory operations. FANUC has operated lights-out factories for twenty-five years, maintaining production without human presence since 2001 (AI Of The Coast). XPENG now runs its IRON humanoid production line with over 80% of core processes automated, targeting mass production by the end of 2026. Xiaomi deployed humanoid robots for full 8-hour shifts at its EV factory, where the machines autonomously sorted center console side covers while catching and correcting their own errors (Fred Lambert from Electrek). Figure reported that its Figure 02 robot completed 10-hour shifts at a BMW facility, loading over 90,000 parts during its deployment (The Rundown AI).

Training data infrastructure. Figure launched an index dataset that gained 69,000 weekly active contributors within two weeks, processing roughly 35 minutes of first-person video every second. JD.com currently utilizes 100,000 internal employees and 500,000 external workers to collect and label robot training data. Chinese embodied AI startups raised 93.5 billion yuan across 322 deals in the first half of 2026, a fivefold increase in funding compared to the previous year as companies shift focus from mechanical movement to software-driven intelligence (TLDR Hardware).

Scaling physical intelligence. Boston Dynamics reduced internal cabling in its new Stretch robot by 81%, allowing a full swerve-drive swap to be completed in one hour rather than a full day. Baidu developed the Apollo World simulator, which generates up to 100,000 kilometers of simulated driving data per day to train autonomous robotaxis. Agility Robotics faces significant unit economic challenges, reporting a gross margin of -151% in 2025 while targeting a bill of materials of $75,000 per unit at a production scale of 1,000 robots annually (Michelle Sun from Core Matter).

• Snowflake is developing CoCo and CoWork bots to handle recurring background tasks and request human intervention only when necessary (TLDR IT).
• NVIDIA introduced VFlash, a vision model that generates geometric predictions and robot trajectories in parallel to improve real-time motion planning (TLDR AI).