Tsinghua Startup Uses 'Uncertain Differential Geometry' to Challenge End-to-End Robot Models

机器之心 · wechat · 2026-08-06

Tsinghua's Xindu Qiyuan team proposes a new embodied AI approach based on 'Uncertainty Theory', offering a white-box alternative to end-to-end models and traditional control theory.

Core Issues: Current probabilistic large models suffer from overconfidence, false precision, and data hunger, causing failures in new environments.
Credibility World Model: Uses 'uncertain differential geometry' to calculate physical 'credible boundaries'. The robot makes optimal decisions within these boundaries without needing massive data training or scene-specific calibration.
Architecture: Features a 7-module transparent pipeline (auditory, tactile, visual, verification, prediction, decision, control) without neural network black boxes, enabling millisecond-level safety checks.
Progress: The fully self-developed system has achieved zero-shot generalization in real-world grasping tasks. The startup is opening a new funding round to expand into industrial applications.

Original post →

Tsinghua Startup Uses 'Uncertain Differential Geometry' to Challenge End-to-End Robot Models

More from Embodied