Baidu's Baizhou Cuts Embodied-AI Data Pipeline Costs: π0.5 Training 2.36x Faster
量子位 · wechat · 2026-11-12
At Baidu World 2026, Baidu Smart Cloud detailed its full-stack embodied-AI infrastructure on the Baizhou platform, spanning data processing, training, deployment and field validation. Baidu says it now serves 70+ embodied-AI companies and holds 29.6% of China's embodied-intelligence cloud market.
The core is data engineering: for million-hour egocentric video, Baizhou uses Lance wide tables for incremental updates and reuse, cuts camera-pose recovery (MegaSaM/Droid-W) GPU hours per video hour by 31–43%, and in OpenPI π0.5 optimization trimmed unused inputs and optimized multi-GPU communication, shrinking a LIBERO epoch from 61.4 to 26 minutes on 8 GPUs (2.36x throughput) with training loss nearly identical to baseline. Customer cases include Xinghaitu (Fast-WAM inference at 87ms/10 steps), Qianxun Intelligence (98.8% effective training time), and D-Robotics voice interaction.
More from Embodied
- GXO to deploy up to 1,000 Apptronik Apollo humanoid robots in warehouses over four years — CyberRobooo · 2026-11-13
- Waymo robotaxi spotted quietly obeying a 'no right turn on red' sign humans often miss — reed · 2026-11-13
- Walden Robotics unveils Sage (LBM 2.0): factory deployments reshape how robots scale — adnothing · 2026-11-13
- Blind Dexterity wins Best Paper at CoRL 2026 workshop: push-resilient humanoid walking using only joint encoders — _krishna_murthy · 2026-11-13
- Bambu Lab blocked a 34x traffic surge that nearly took its cloud down, meme follows — ngxson · 2026-11-13
- FLASH-ALOHA open-sources leader-arm-only teleop hardware, CAD, and code — philfung · 2026-11-13