2024.09.24 - #8 - World models (Wayve, 1X, World Labs), MaskBEV, RockChip micro-NPU, M2-Mapping

[changh95]

Sponsor

• VRCREW 엔지니어 채용
  • 회사 페이지 링크
  • 채용 공고 링크

• VRCREW 라이다 & 3D 스캐너 중고판매

Industry trend

• Wayve neural simulator: Ghost Gym (링크)
  • 자율주행을 위한 뉴럴 시뮬레이터 (실제 센서로 취득한 데이터로 설계한 world model)

• 1X world model (i.e. virtual simulator from real sensor data) for robot manipulation (링크)

• World labs, Fei-Fei Li 교수님 스타트업 (링크)
  • LLM이 아닌 Large world model 연구개발
  • $230M (3000억 원) 초기투자

• Microsoft, Blackrock 등등 AI 데이터센터를 위한 1천억 달러 (130조 원) 투자 링크

Lecture series

• Stanford CS149 Parallel computing
  • CPU 멀티코어 프로그래밍, GPU 병렬처리 (CUDA)

[rubidiumchain]

(Pilippeshin)

Interesting Paper

MaskBEV: Towards A Unified Framework for BEV Detection and Map Segmentation

This paper proposes a unified and general multimodal multi-task learning (MTL) paradigm. MaskBEV completes multi-task 3D perception based on bird’s eye view (BEV) representation in a shared Transformer decoder.

Model. Multimodal feature encoder can be BEVFusion or UniTR

Training. Different from the training strategy of the general architecture that separates the BEV encoders for different tasks,
our MaskBEV jointly trains 3D detection and BEV map segmentation tasks in the unified encoder-decoder framework.
We trained our model on 8 NVIDIA A800 GPUs by AdamW optimize

NVLink vs UALink vs CXL

NVLink 및 NVLink Switch

차세대 AI 가속기 연결 표준 'UALink' 개발 위해 NVIDIA 제외한 AMD, Intel 등 대형 IT 기업들이 연합

컴퓨트 익스프레스 링크(CXL)™

Rock-chips's New RV1103, RV1109, RV1126 /w MicroNPU

RV11 Series

RV1103-0.5Tops NPU
RV1109-1.2Tops NPU
RV1126-2.0Tops NPU

Luckfox Pico Mini A RV1103 Linux Micro Development Board

Brief Datasheet of RV1109

The RV1126 and RV1109 have the same functional interface, with the exception that the RV1126 is a quad-core and the RV1109 is a dual-core, and the RV1126 is slightly more powerful. The RV1126 has a 1400-pixel ISP processor with 2TOPS of NPU arithmetic, whereas the RV1109 has a 5MP ISP processor with 1.2TOPS of NPU arithmetic. Both are P2P designs, easy to replace, and can effectively reduce development costs by forming a high and low-end product line.

[U-AMC]

Papers

M2Mapping
Neural Surface Reconstruction and Rendering for LiDAR-Visual Systems

HK_MEMS_Dataset
HK-MEMS, a MEMS LiDAR dataset on urban tunnels and dynamic scenarios

Industry

텐센트에서 새로운 로봇 발표...!
"The Five"

IdsDVjh9fMaNNSes.mp4

PQ_WBsOD0-RZK0Mg.mp4

Blogs

Unitree Go2
Unitree Go2를 두대나 산 이유에 대하여... Personal review & demos