ChanganVR
commented
5 months ago Hi @Yuanzizizi, it is probably not surprising that you see degraded performance when testing the algorithm in unseen configurations due to distribution mismatch. The most straightforward way to fix this would be to incorporate more scenarios and randomize configurations as much as possible during training so that the model can generalize to a wider set of scenarios at test time.
We are interested in understanding more about the generalization capabilities of your algorithm. Specifically, we have noticed that when testing in scenarios where the robot's position or the strategies of other humans have been altered, the performance seems to degrade significantly. Could you provide insights into why such changes in the environment or agent behaviors affect the effectiveness of the algorithm? We are curious to know if there are any strategies or techniques employed to enhance the algorithm's adaptability to varying scenarios and environmental conditions. Any information regarding the algorithm's robustness and its ability to generalize across different scenarios would be greatly appreciated.