Transfer from Imprecise and Abstract Models to Autonomous Technologies (TIAMAT)*
Abstract
*Formerly Learning Autonomy in Synthetic Environments (LASE) The Transfer from Imprecise and Abstract Models to Autonomous Technologies (TIAMAT) program will develop techniques to robustly transfer learned autonomy from fast abstract simulations to autonomous platforms in real-world environments. The autonomy levels of unmanned systems of today are limited because the modeling and simulation (M&S) training environments do not account for the data domain shift common when translating M&S outcomes to the real world - this phenomenon is sometimes referred to as the sim2real gap. The TIAMAT approach will integrate symbolic structures with neural structures to more realistically and robustly transfer learned autonomy. TIAMAT will enable the use of fast abstract simulations by anchoring the learning and transfer of autonomy on semantically consistent components shared across simulations and real environments, so-called "semantic anchors". For TIAMAT, semantic anchors of particular importance include those militarily-relevant phenomena that remain consistent in the source and target environments, for example, mission objectives, special instructions, subject matter expert guidance, rules of engagement, and the laws of physics. Autonomy transfer using semantic anchors will reduce the complexity of the autonomy learning and transfer problems to the comparatively simpler points of reference in the anchored representation. If successful, TIAMAT transfer of M&S-based learning will enable more rapid and robust training and deployment of autonomous systems at higher levels of autonomy.
Document Details
- Document Type
- Accomplishment
- Publication Date
- Oct 01, 2025
- Source ID
- e5498b90ede85440bbb54db17188005f