Table 2.

Estimation performance (mean of RMSEs) for each RNN model over three augmented test datasets (unit: °).

(a) Results for the test dataset with applied virtual gyroscope bias
	Attitude	Heading	Avg. Improvement
Original	9.27	12.09	-
Rotation	6.60	8.25	30.3%
Bias	7.33	9.47	21.3%
Noise	7.98	11.02	11.4%
Rotation and Bias	6.63	8.64	28.5%
Rotation and Noise	6.60	9.04	27.0%
Bias and Noise	7.37	9.45	21.2%
All	6.00	7.85	35.2%
(b) Results for the test dataset with applied virtual noise
	Attitude	Heading	Avg. Improvement
Original	9.27	12.11	-
Rotation	6.59	8.26	30.4%
Bias	7.32	9.49	21.3%
Noise	7.98	11.03	11.4%
Rotation and Bias	6.63	8.65	28.5%
Rotation and Noise	6.60	9.05	27.0%
Bias and Noise	7.36	9.47	21.2%
All	5.99	7.85	35.2%
(c) Results for the test dataset with applied virtual rotation
	Attitude	Heading	Avg. Improvement
Original	30.43	38.57	-
Rotation	13.00	14.63	59.7%
Bias	25.98	33.79	13.5%
Noise	27.24	35.99	8.6%
Rotation and Bias	12.36	15.77	59.3%
Rotation and Noise	11.58	14.31	62.3%
Bias and Noise	25.81	31.95	16.2%
All	11.60	14.45	62.2%
(d) Results for the dataset from [39]
	Attitude	Heading	Avg. Improvement
Original	31.49	54.77	-
Rotation	26.40	31.24	29.6%
Bias	30.90	45.16	9.71%
Noise	33.72	50.82	0.05%
Rotation and Bias	24.77	32.69	30.8%
Rotation and Noise	23.74	34.71	30.6%
Bias and Noise	34.36	49.01	0.69%
All	24.63	31.28	32.3%