We show both simulation and real scenarios of all six tasks:

Go through, Crawl, Distinguish, Go avoid, Unload and Go to

We here show results in different sim2real training paradigms and failure case analysis.

1. Comparing Model 3 with Models 1 & 2: Through fine-tuning, it is observed that the average motion length is significantly reduced, aligning closely with the performance of models trained exclusively on real data. This indicates that the incorporation of real data enhances the model's ability to recognize objects with greater precision and speed, thereby mitigating some discrepancies between real and simulated data.

2. Comparing Model 2 with Model 1: It is evident that with a reduced amount of simulation data, the model's proficiency in yaw control is compromised, leading to operational failures. This underscores the importance of an adequate dataset size for the effective acquisition of VLA skills.

3. Comparing Model 3 with Models 1 & 4: Upon examination, it becomes clear that utilizing a co-training approach, the model manages to learn a gait pattern in the simulated environment that closely resembles that of real-world data. This demonstrates the effectiveness of simulation-based learning, as the skills acquired can be effectively transferred to real-world scenarios.