Table 2.

Average success rate (%) and standard deviation obtained by different methods in three testing environments.

Env	Laser-Only	Depth-Only	Multi	Multi-AT	Multi-BF	MDRLAT
Env1	67.6 $\pm$ 4.8	69.2 $\pm$ 3.5	76.8 $\pm$ 4.1	86.0 $\pm$ 4.2	89.6 $\pm$ 1.5	94.4 $\pm$ 2.3
Env2	61.6 $\pm$ 4.6	55.2 $\pm$ 4.8	73.6 $\pm$ 5.3	78.8 $\pm$ 2.7	81.2 $\pm$ 3.0	87.2 $\pm$ 2.4
Env3	64.8 $\pm$ 3.2	63.2 $\pm$ 2.7	70.8 $\pm$ 4.1	78.4 $\pm$ 5.0	79.6 $\pm$ 3.4	83.6 $\pm$ 3.2

Laser-only, directly feed the feature representation extracted from 2D laser range findings into D3QN; depth-only, directly feed the feature representation extracted from depth images into D3QN; multi, directly concatenate the feature representations extracted from different sensor modalities; multi-AT, adopt the auxiliary task module in the multi method; multi-BF, adopt BF module in the multi method; MDRLAT, our proposed method.