01
Joint-masked layering
Locomotion, speech-gesture and face combine via per-joint masking so a character can walk and talk at the same time without one motion fighting the other.
In development · AI & Avatars · Engine
Real-time procedural avatar engine. Context-tagged motion knowledge base distilled from teacher models, skeleton-agnostic retargeting, joint-masked layering of locomotion, speech-gesture and face. Powers walk-and-talk for embodied AI.
01 · Overview
Bvatar is Vrtron's internal avatar engine: a real-time, procedural system for embodied AI characters that can locomote, gesture, look, and speak in plausible combination.
At its heart is a context-tagged motion knowledge base built by distilling teacher models, plus a skeleton-agnostic retargeting layer so the same motions drive different character rigs without per-character authoring. Locomotion, speech-gesture and face are independent layers that combine through joint-masked blending.
Bvatar powers walk-and-talk behaviour in TOW's NPC framework and the Borromini Kiosk, and is the reference runtime for the Vrtron Avatar product (separately developed for client work).
02 · Highlights
01
Locomotion, speech-gesture and face combine via per-joint masking so a character can walk and talk at the same time without one motion fighting the other.
02
ZeroEGGS / SMPL-X / Ubisoft BVH round-trips handled in-engine. Round-trip MAE ~0.69 cm at body scale.
03
Motions are tagged with the context that produced them (intent, prosody, scene), so the engine can sample the right gesture for the right moment.
04
Large generative teacher models capture the motion space; Bvatar runs the distilled student fast enough for real-time on consumer GPUs.
03 · Stack