NVIDIA's VFlash: a generalist parallel decoder for structured Physical AI outputs
ZhidingYu · x · 2026-11-06
Part 3 of a thread by NVIDIA researcher Zhiding Yu on the requirements Physical AI places on decoders and the design of VFlash.
Characteristics of Physical AI tasks:
1) Rich output types: outputs span both text (thinking, sub-task planning) and various visual primitives;
2) Structured outputs: points, boxes, and trajectories have deterministic structures;
3) Versatile requirements: NAVSIM needs 8 future waypoints over 4s at 2Hz, Waymo E2E needs 20 over 5s, and Alpamayo predicts 64 over 6.4s — so a generalist decoder needs on-demand variable-length support.
VFlash extends box-aligned decoding from LocateAnything to richer Physical AI output types with systematic parallel decoding designs, introducing a structured tool call to declare the length of variable-size primitives.
More from Embodied
- GPT-6 Astra tops a 16-benchmark robot index, yet went 0/11 on real plug insertions — AIFlow_ML · 2026-11-06
- Giving Claude a cutting-edge robot body just so it can spoon-feed me — repligate · 2026-11-06
- moonlake previews research on scaling robot learning with simulation in the RL era — du_yilun · 2026-11-06
- Nearly 100 testers praise Sphenomenon MR demo that turns a whole house into voxels — Scobleizer · 2026-11-06
- Spectral Skills lands on real robots: composing motion skills via energy functions — daibond_alpha · 2026-11-06
- Developer gives Claude Desktop a body and sensors via MCP and Raspberry Pi — jfrankie · 2026-11-06