| 項次 | 料號 | 數量 | 單位 | 需用日 | 小計 |
| | 品名規格 | | | 單價 | |
| 規格 | |
| 10 |
|
1 |
CS |
20261211 |
|
| | 委託"用於AI晶片推論之視覺語言模型機器人動作決策技術服務" | | |
________________ |
________________ |
| | Here are the key components of the module: Digital-twin pick and place module: on-premises Intrinsic Core; the whol e module runs locally without internet connection 1. Identify the vision-based decision points of a physics-simulated tool -tidying cell holding at least five different tools (including a pair of scissors, a screwdriver and an adjustable wrench) represented by photor ealistic 3D models, initially placed at random, each to be returned to i ts printed outline at a specified 2D pose (position and orientation), ob served through a simulated wrist camera, with partially overlapping obje cts at which object identity, occlusion order or outline occupancy deter mines the next robot action 2. For the decision points identified in 1., design a decision scheme in which scene facts from a segmentation model and depth geometry constrai n the VLM, such that an object lying under another is never moved first (models developed in mainland China are not allowed) 3. Determine the memory usage from the image-token count, scene-fact len gth, answer length and KV cache size required by each decision query for given object count and history length, so that the memory allocated for inference on AI chips can be optimized effectively Performance requirements: In at least 20 test runs, each starting from a different random placement of the tools in 1. and together covering at least three natural-language task instructions that differ in the requi red order of placement, achieve a decision accuracy of at least 90% as j udged by an oracle policy, an average decision latency of no more than 6 seconds per decision, and a KV cache size of no more than 4 GB per deci sion query Technical report: Describe the techniques developed in 1. to 3., and com pare the scheme with image-only decision (the same VLM deciding from the photo alone, without 2.) and with an oracle policy in terms of task suc cess rate, decision latency and memory footprint Source code: Provide the source code of 1. to 3. and of the tests in the technical report, with build and run instructions, so that the reported results can be reproduced on the on premises Intrinsic Core
|
|