Computer Vision News - December 2021

8 Robot Learning Research RECON is also proven to be stable amidst previously unseen obstacles and weather conditions. This is warranted by the invariance of the learned representation to such factors. To test this, the authors made RECON first explore a new “junkyard” to learn to reach a goal image containing a blue dumpster, and then they evaluated the learned goal-reaching policy when presented with previously unseen obstacles or lighting conditions. In the figure above, under Novel Obstacles the trajectory slightly changes to avoid the obstacles, while under varying lighting conditions the trajectory is not affected. Finally, some ablation studies were also led to ensure that the building blocks of the RECON method (the topological graph for memory, rollouts to sampled goals, the information bottleneck) are fundamental to its robustness and performance. This was confirmed by experiments showing that the full algorithm substantially improves on the other variants’ timings (1.58-4.58 minutes for exploration time, 2.9-11.4 seconds for navigation time). formance using Success weighed by Completion ount the agen ’s dy amics [58]. We show quantita- ctories of RECON and the baselines in Fig. 4. Figure 5: Exploring non-stationary environ- ments: The learned repres ntation an topol g- ical graph is robust to visual distractors, enabling reliable navigation to the oal under novel obsta- cles (c–e) and appearance changes (f–h) . ing in- ine and on- ver and 5 m ely ing on nvi- ra- re- uire on- ess de- can ach RE- ble act topological map fr m experience in the target viously-seen states. The other baselines are unsuc-