{"id":2784,"date":"2023-02-02T17:11:53","date_gmt":"2023-02-02T07:11:53","guid":{"rendered":"https:\/\/www.cognav.net\/?p=2784"},"modified":"2023-02-02T17:11:53","modified_gmt":"2023-02-02T07:11:53","slug":"do-artificial-intelligence-ai-agents-learn-to-build-internal-spatial-representations-or-mental-maps-of-their-environment-as-a-natural-consequence-of-learning-to-navigate","status":"publish","type":"post","link":"https:\/\/braininspirednavigation.com\/?p=2784","title":{"rendered":"Do artificial intelligence (AI) agents learn to build internal spatial representations (or \u2018mental\u2019 maps) of their environment as a natural consequence of learning to navigate?"},"content":{"rendered":"<p style=\"text-align: justify;\">Erik Wijmans, Manolis Savva, Irfan Essa, Stefan Lee, Ari S. Morcos, Dhruv Batra.\u00a0<strong><a href=\"https:\/\/arxiv.org\/abs\/2301.13261\">Emergence of Maps in the Memories of Blind Navigation Agents<\/a><\/strong>. arXiv:2301.13261 [cs.AI], 2023.\u00a0<\/p>\n<p style=\"text-align: justify;\">Abstact<br \/>\n&#8220;Animal navigation research posits that organisms build and maintain internal spatial representations, or maps, of their environment. <strong><span style=\"color: #ff0000;\">We ask if machines &#8212; specifically, artificial intelligence (AI) navigation agents &#8212; also build implicit (or &#8216;mental&#8217;) maps<\/span><\/strong>. A positive answer to this question would (a) explain the surprising phenomenon in recent literature of ostensibly <span style=\"color: #ff0000;\"><strong>map-free neural-networks achieving strong performance<\/strong><\/span>, and (b) strengthen the evidence of <span style=\"color: #ff0000;\"><strong>mapping as a fundamental mechanism for navigation by intelligent embodied agents<\/strong><\/span>, whether they be biological or artificial. Unlike animal navigation, we can judiciously design the agent&#8217;s perceptual system and control the learning paradigm to nullify alternative navigation mechanisms. Specifically, <strong><span style=\"color: #ff0000;\">we train &#8216;blind&#8217; agents &#8212; with sensing limited to only egomotion and no other sensing of any kind &#8212; to perform PointGoal navigation (&#8216;go to \u0394 x, \u0394 y&#8217;) via reinforcement learning<\/span><\/strong>. Our agents are composed of navigation-agnostic components (fully-connected and recurrent neural networks), and our experimental setup provides no inductive bias towards mapping. Despite these harsh conditions, <strong><span style=\"color: #ff0000;\">we find that blind agents are<\/span><\/strong> (1) surprisingly <strong><span style=\"color: #ff0000;\">effective navigators in new environments (~95% success)<\/span><\/strong>; (2) they<strong><span style=\"color: #ff0000;\"> utilize memory over long horizons<\/span> <\/strong>(remembering ~1,000 steps of past experience in an episode); (3) this <strong><span style=\"color: #ff0000;\">memory enables them to exhibit intelligent behavior<\/span><\/strong> (following walls, detecting collisions, taking shortcuts); (4) there is <strong><span style=\"color: #ff0000;\">emergence of maps and collision detection neurons in the representations of the environment built by a blind agent as it navigates<\/span><\/strong>; and (5) <strong><span style=\"color: #ff0000;\">the emergent maps are selective and task dependent<\/span><\/strong> (e.g. the agent &#8216;forgets&#8217; exploratory detours). Overall, this paper presents no new techniques for the AI audience, but a surprising finding, an insight, and an explanation.&#8221;<\/p>\n<p style=\"text-align: justify;\">Erik Wijmans, Manolis Savva, Irfan Essa, Stefan Lee, Ari S. Morcos, Dhruv Batra.\u00a0<strong><a href=\"https:\/\/arxiv.org\/abs\/2301.13261\">Emergence of Maps in the Memories of Blind Navigation Agents<\/a><\/strong>. arXiv:2301.13261 [cs.AI], 2023.\u00a0<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Erik Wijmans, Manolis Savva, Irfan Essa, Stefan Lee, Ari S. Morcos, Dhruv Batra.\u00a0Emergence of Maps in the Memories of Blind Navigation Agents. arXiv:2301.13261 [cs.AI], 2023.\u00a0 Abstact &#8220;Animal navigation research posits that organisms build and maintain internal spatial representations, or maps, of their environment. We ask if machines &#8212; specifically, artificial intelligence (AI) navigation agents &#8212; [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[96,376,519],"tags":[1176,1175,648],"_links":{"self":[{"href":"https:\/\/braininspirednavigation.com\/index.php?rest_route=\/wp\/v2\/posts\/2784"}],"collection":[{"href":"https:\/\/braininspirednavigation.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/braininspirednavigation.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/braininspirednavigation.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/braininspirednavigation.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2784"}],"version-history":[{"count":1,"href":"https:\/\/braininspirednavigation.com\/index.php?rest_route=\/wp\/v2\/posts\/2784\/revisions"}],"predecessor-version":[{"id":2785,"href":"https:\/\/braininspirednavigation.com\/index.php?rest_route=\/wp\/v2\/posts\/2784\/revisions\/2785"}],"wp:attachment":[{"href":"https:\/\/braininspirednavigation.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2784"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/braininspirednavigation.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2784"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/braininspirednavigation.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2784"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}