Draft of talk slides

slides.md

@@ -84,43 +84,113 @@ layout: two-cols-header
 <ul>
 <v-click at="+1"><li>Children (and adults) create playful goals</li></v-click>
 
-<v-click at="+2"><li>These goals help us learn how to search and solve problems
-<span class="text-xs">[Chu & Schulz, 2020; Chu et al., 2024]</span></li></v-click>
+<v-click at="12"><li>These goals help us learn how to structure problem spaces and find solutions
+<span class="text-xs"><br>[Chu & Schulz, 2020; Molinaro & Collins, 2023; Chu et al., 2024]</span></li></v-click>
 </ul>
 </div> 
 </div>     
 
 <div>
     <div class="flex flex-row justify-center">
         <figure class="image is-square is-inline-block">
-            <v-click at="[10, 11]"><img src="/images/logan_truck.jpg" width="160"></v-click>
-        </figure>
-        <figure class="image is-square is-inline-block">
-            <v-click at="[11, 12]"><video width="240"controls>
-                <source src="/videos/jordan_bird_trimmed.mp4" type="video/mp4">
-            </video></v-click>
+            <v-switch>
+                <template #1><img src="/images/logan_truck.jpg" width="240"></template>
+                <template #2><SlidevVideo width="320" controls>
+                    <source src="/videos/jordan_bird_trimmed.mp4" type="video/mp4">
+                </SlidevVideo></template>
+            </v-switch>
         </figure>
     </div>
 </div>
 
 </div>
+
+
+---
+layout: fact
+---
 
+## Core idea:
+<br>
+If we want to develop agents that accomplish diverse tasks across different environments, 
 
+we need agents that can propose and pursue rich, complex, and creative goals.
+<br>
+<br>
+<br>
+<span class="text-xs">\[Ouedeyer et al., 2007; Colas et al., 2023]</span>
 
-<v-click at="+1"><h3></h3></v-click>
-<div class="ns-c-tight">
-<v-clicks at="+2">
+---
+
+# Desiderata for Goal Representations
+- Abstraction (abstract goals, abstracting goal components)
+- Temporal extension
+- Compositionality
+- Grounding
+<br><br><br>
+- These properties likely neither necessary nor sufficient
+
+---
 
+# Surveyed Goal Representation Approaches
+<v-clicks depth="2">
 
+- Implicit (directly encoded as reward functions)
+- Goal states (e.g. target manipulator positions)
+- Image-based observations 
+- Natural language:
+    - Environment-provided
+    - Procedurally generated from minimal grammars
+    - Language-based exploration
+    - Multimodal models
+- Represented as programs? 
 
 </v-clicks>
-</div>
 
+<!-- <div class="text-xl absolute top-1/2 bg-slate-200 border-2 border-black py-6" v-click>
+Prevalent non-language approaches facilitate grounding at the expense of other desiderata
+<br><br>
+Language (and programs) offer benefits at the cost of grounding complexity
+</div> -->
+
+
+---
+layout: fact
+---
+
+## Key takeaway:
+<br>
+Prevalent non-language approaches facilitate grounding at the expense of other desiderata
+<br><br>
+Language (and programs) offer benefits at the cost of grounding complexity
+
+---
+
+# Example argument: compositionality
+- **Reward functions:** Compose mathematically, not semantically
+- **Observations:** Composing images to represent general properties is hard
+- **Language:** Inherently compositional, but grounding is hard 
+    - SuccessVQA near chance on held-out goals \[Du et al., 2023a]
+    - Hill et al. \[2019] generalize to held-out objects, but not negations
+- **Programs:** Compose by default, as defined by their grammar
+    - The structured LTL-based approach of Leon et al. \[2022] composes negation succesfully 
+
+
+---
+# Takeaways
+<v-clicks depth="2">
+
+- **Takeaway 1:** We need agents that can propose and pursue rich, complex, and creative goals.
+- **Takeaway 2:** This requires richer goal representations and developing methods to ground them.
+<br><br>
 
-<!-- ::left::
+## Questions
 
-- Specifying goals as intended behaviors:
+- Where do goals (and their representations) fall on the agent-environment boundary?
+- What important desiderata are we missing?
+- How do temporally extended goals play with the Markov assumption?
+- Can program-based goals scale to diverse environments?
+- What can building agents that propose and pursue rich goals teach us about human goal-setting?
 
-:::right:::
 
-- Children's ability to propose goals in play.  -->
+</v-clicks>