Table 18:

Complexity results for datasets in the robotics domain. U: unimodal models, M: multimodal fusion paradigms, O: optimization objectives, T: training structures.

	Dataset	MuJoCo Push
	Metric	Epochs trained	Training time (s)	Training params (M)	Training peak memory (MB)	Inference time (s)	Inference params (M)
U	Unimodal $(i)$ Unimodal $(f)$ Unimodal $(p)$ Unimodal $(c)$	20 20 20 20	738±133 288±39 252±6 372±64	3.88 3.33 3.33 3.33	3607±1 3595±2 3594±1 3594±1	3.46±0.02 0.91±0.08 0.87±0.04 0.86±0.04	3.88 3.33 3.33 3.33
M	EF LF-LSTM TF-LSTM [179] MulT [156]	20 20 20 20	815±34 856±46 1914±31 4792±62	3.92 1.90 23.5 14.6	3654±1 3636±1 4530±9 6530±16	4.44±0.55 4.32±0.45 7.75±0.12 22.4±0.28	3.92 1.90 23.5 14.6

	Dataset	Vision&Touch
	Metric	Epochs trained	Training time (s)	Training params (M)	Training peak memory (MB)	Inference time (s)	Inference params (M)
U	Unimodal $(i)$ Unimodal $(f)$ Unimodal $(p)$	15 1 5 15	2633 2185 2514	1.00 0.13 0.08	5530 2426 2389	63.9 51.6 59.5	1.00 0.13 0.08
M	LF Sensor Fusion [91] LRTF [106]	15 50 35	2672 11604 8366	1.20 1.10 1.09	5572 4467 4987	64.4 62.6 64.4	1.20 1.10 1.09
O	RefNet [135]	15	3819	135	6067	65.0	1.20