Inside Roblox's Bet on World Models

KẾT LUẬN

Roblox dang phat trien kien truc lai Roblox Reality ket hop game engine truyen thong voi video world model Super Upsampler: engine lam nguon chan ly (state, physics, rules, multiplayer), model chi upsample frame tho thanh photorealistic — dat duoc do chan thuc thi giac voi chi phi thap ma van giu tinh nhat quan va kha nang choi nhieu nguoi.

LUẬN ĐIỂM CHÍNH

• Video world model thuan tuy khong the thay the game engine — model thieu bo nho ben vung (persistence), khong co quy tac nhat quan, khong the bien player input thanh ket qua chinh xac. Khi quay camera di roi quay lai, the gioi da thay doi.
• Game engine thuan tuy khong the dat photorealism voi chi phi hop ly — render tat ca chi tiet (texture phan giai cao, anh sang phuc tap, hieu ung vat li) qua dat do, bat kha thi voi studio nho va thiet bi pho thong.
• Kien truc lai chia cong viec ro rang: engine render frame tho (dung hinh dang, vi tri, chuyen dong, depth map) -> model upsample thanh anh chan thuc (texture, anh sang, be mat, hieu ung thoi tiet) — model khong duoc phep thay doi vi tri hay cau truc cua bat ky doi tuong nao.
• Bon van de mo truoc khi ship: (1) latency — can tu ~5 giay xuong ~30ms bang self-forcing; (2) consistency dai han — tranh the gioi troi dan qua hang phut choi; (3) multiplayer — quan ly 20,000 nguoi choi voi 20,000 trang thai rieng biet, khong chi la ve dam dong; (4) creator control — lam sao de nguoi sang tao dieu khien duoc model thay vi model tu do sinh noi dung.

LUẬN ĐIỂM PHỤ

• Data model (ban ghi co cau truc cua moi doi tuong trong the gioi) la chia khoa chong troi — model chi quyet dinh ve ngoai, engine la authority duy nhat ve nhung gi ton tai. State luu trong data model tren server, pixel luu trong video latents va duoc tai sinh lien tuc.
• Self-forcing (tu Morpheus AI) bien model offline cham thanh autoregressive, moi frame duoc sinh dieu kien boi cac frame truoc do -> stream lien tuc thay vi batch. Ket hop voi KV cache nen, GPU H200/B200 tai edge data centers gan nguoi choi.
• Creator control thong qua game cartridge harness (tu Lucid AI) — boc logic xac dinh cua game engine quanh video world model, cho phep the gioi trong nhu sinh ra tu AI nhung van hanh theo quy tac co dinh do nguoi sang tao thiet lap.

LUẬN CỨ & VÍ DỤ

• Dan chung ve thieu persistence: neu quay camera di roi quay lai, trong hau het video model the gioi se khong con giong nhu truoc.
• Dan chung ve khac biet pixels vs simulation: model co the ve dam dong 500 nguoi, nhung do chi la pixel, khong phai 500 nguoi choi rieng biet voi vi tri, inventory va hanh dong cua chinh ho.
• Dan chung ve model khong hieu vat ly: model co the ve mot chai nuoc hoan hao nhung khong biet no la cai chai, khong biet no mo ra, dung nuoc, hay roi xuong khi buong tay — pixels khong phai la su hieu biet.
• Dan chung ve phan cong engine-model: engine render frame plain dung hinh hoc nhung texture don gian -> Super Upsampler them gloss tren son uot, glare tren kinh chan gio, mat duong nham, bui nuoc ban sau xe — khong bao gio di chuyen xe hay thay doi khoang cach.

Nguồn: ByteByteGo | Phân tích bởi AI