欢迎大家点赞、关注集群式无人机板块(今年会全面完善仿真环境、算法、真实部署等多个部分)～

1. Multi-robot collaborative perception

1.1 conception

• Collaborative perception learns how to share information among multiple robots to perceive the environment better than individually done.

1.2 Motivation

• Modern perception systems of embodied robots are known to be sensitive to occlusions and sensor degradations or failures, and lack the capability of long perceiving range.
  

1.3 Category

• N robots - 1 targets
• N robots - N targets
• N robots - M targets
  

1.4 Framework

(1) process sensor data, (2) broadcasting it, (3) incorporation it, (4) generate results.

• Early Fusion: The vehicles will directly transmit the raw point clouds to other collaborators.
• Intermediate Fusion: The collaborators will extract intermediate features using a neural feature extractor and broadcast the compressed feature to other collaborators.
• Late Fusion: Each vehicle detects 3D objects utilizing its own sensor observations and delivers the predictions to others. Then the receiver applies Nonmaximum suppression to produce the final outputs.
  

• １－原始传感器数据：可以最大限度地减少信息损失，但它们需要更多的带宽，此外，接收车辆需要处理所有收到的额外的传感器数据，这可能使它无法满足实时推理的要求。
  ２－中间特征：深层网络中的中间表示可以轻松压缩，同时保留下游任务的重要信息．并且计算开销低，因为来自其他智能体的传感器数据已经被预处理。
  ３－任务输出：可以最大程度降低传输带宽开销，因为只有少数数字需要被广播，但可能会失去有价值的场景内容和不确定性信息，而这些信息对于更好地融合信息可能是非常重要的。

1.5 challenge

• When does robots communicate, with whom, and with what?
• How to imcorporate?

---

---

2. Multi-robot collaborative raw data, encoding, decoding

year	Method	Venue	sensor	dataset	eoncoding	decoding	tasks	loss
2020	Who2com	ICRA	LiDAR
2020	when2com	CVPR	LiDAR
2020	V2VNet	ECCV	LiDAR
2021	DiscoNet	NeurIPS

3. Multi-robot collaborative communication

year	Method	Venue	sensor	graph construction	compress	who	when	what	error
2020	Who2com	ICRA	LiDAR	Handshake-based sparse graph	-	Y	-	Full feature map	-
2020	when2com	CVPR	LiDAR	Handshake-based sparse graph	-	Y	Y	Full feature map	-
2020	V2VNet	ECCV	LiDAR	Fully connected graph	-	-	-	Full feature map	-
2021	DiscoNet	NeurIPS	LiDAR	Fully connected graph	-	-	-	Full feature map	-

4. Multi-robot collaborative perception fusing

year	Method	Venue	sensor	pose	fusing	delay	asynchrony	view	error
2020	Who2com	ICRA	LiDAR
2020	when2com	CVPR	LiDAR
2020	V2VNet	ECCV	LiDAR
2021	DiscoNet	NeurIPS	LiDAR