رونوشت 🔗
1
00:00:03,000 --> 00:00:08,520
Evan: Welcome to Daily Paper Cast, bringing you the latest breakthroughs in AI research.
2
00:00:08,568 --> 00:00:16,908
Ashley: Today's paper is from Hugging Face's daily paper list of October 14, 2026, and has received 58 upvotes.
3
00:00:16,968 --> 00:00:23,708
Evan: The paper we're discussing is titled 'Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes.'
4
00:00:23,760 --> 00:00:30,080
Ashley: The first two authors are Yu Wang and Rui Dai, and the corresponding author is Xiangxiang Chu.
5
00:00:30,080 --> 00:00:35,200
They are affiliated with Wuhan University and the DreamX Team at Alibaba Group.
6
00:00:35,316 --> 00:00:40,476
Evan: So, Ashley, what is the main challenge that this paper addresses?
7
00:00:40,536 --> 00:00:48,936
Ashley: The core challenge tackled by this paper is grounding objects in large-scale 3D urban scenes using natural language descriptions.
8
00:00:48,936 --> 00:00:57,496
This capability is essential for embodied intelligence applications like autonomous driving, robotic navigation, and aerial inspection.
9
00:00:57,496 --> 00:01:08,096
Indoor 3D grounding has been extensively studied, but city-scale 3D grounding remains underexplored due to the vast spatial extent and complexity of urban scenes.
10
00:01:08,160 --> 00:01:09,510
Evan: That's interesting.
11
00:01:09,510 --> 00:01:14,600
Why is city-scale grounding particularly difficult compared to indoor settings?
12
00:01:14,664 --> 00:01:25,224
Ashley: Outdoor urban environments can contain tens of millions of points per scene and require reasoning over long-range spatial cues and abstract referring expressions.
13
00:01:25,224 --> 00:01:51,524
These environments feature repetitive elements like vehicles and buildings, and queries may describe objects via landmarks or spatial relationships, such as 'a white car parked next to the orange car to the left of the West Midlands Police Custody Suite.' Moreover, existing methods designed for indoor scenes typically downsample large point clouds, losing fine-grained geometry and leading to suboptimal performance.
14
00:01:51,576 --> 00:01:57,736
Evan: But how do current city-scale grounding methods try to tackle this issue?
15
00:01:57,792 --> 00:02:12,592
Ashley: Recent methods for city-scale grounding often rely on a candidate-ranking approach, which involves precomputing instance proposals from annotations or using instance segmentation pipelines and then scoring these candidates based on text-instance matching.
16
00:02:12,592 --> 00:02:18,422
Though this reduces the input scale, it's tightly coupled to the quality of these precomputed candidates.
17
00:02:18,422 --> 00:02:24,602
The accuracy depends on the proposal quality, and the inference cost grows with the number of candidates scored.
18
00:02:24,602 --> 00:02:32,172
Importantly, this approach treats grounding as a static retrieval problem rather than a dynamic and exploratory process.
19
00:02:32,232 --> 00:02:36,892
Evan:
Load full transcript
So how does Urbex propose to solve this problem differently?
20
00:02:37,056 --> 00:02:43,116
Ashley: Urbex reformulates city-scale 3D grounding as an active spatial search problem.
21
00:02:43,116 --> 00:02:52,216
Instead of relying on precomputed candidates, Urbex represents a city-scale 3D scene as an interactive multi-view environment.
22
00:02:52,216 --> 00:03:05,936
A vision-language model (VLM) agent actively queries landmarks, zooms into local regions, renders oblique views, and commits to grounded bounding boxes based on accumulated evidence.
23
00:03:06,000 --> 00:03:08,770
Evan: That sounds like a more dynamic approach.
24
00:03:08,770 --> 00:03:11,380
How is this search process optimized?
25
00:03:11,424 --> 00:03:32,104
Ashley: The optimization uses reinforcement learning on the CityRefer dataset, which includes using smooth localization rewards when the overlap with the target is initially sparse and lightweight shaping rewards to guide intermediate behaviors like zooming toward target regions.
26
00:03:32,104 --> 00:03:41,604
This multi-step decision process helps Urbex optimize both final localization accuracy and the agent's search policy.
27
00:03:41,664 --> 00:03:44,744
Evan: And the results from deploying Urbex?
28
00:03:44,808 --> 00:04:00,978
Ashley: Experiments show that Urbex achieves the best Hit@1 and Intersection over Union (IoU) on CityRefer, supports candidate-free grounding without ground-truth boxes, and is substantially faster than the strongest city-scale ranking baseline.
29
00:04:00,978 --> 00:04:08,708
Additionally, zero-shot results on STPLS3D-Refer indicate improved cross-scene robustness.
30
00:04:08,760 --> 00:04:10,310
Evan: Sounds promising.
31
00:04:10,310 --> 00:04:12,590
That wraps up the Introduction section.
32
00:04:12,590 --> 00:04:16,220
Next, we'll dig into the methods that make this possible.
33
00:04:17,534 --> 00:04:23,644
Evan: Alright Ashley, let's delve deeper into the innovative methods proposed by Urbex.
34
00:04:23,644 --> 00:04:26,174
Can you walk us through the core components?
35
00:04:26,234 --> 00:04:27,334
Ashley: Sure, Evan.
36
00:04:27,334 --> 00:04:34,994
Urbex formulates city-scale 3D grounding as an active spatial search problem rather than relying on static candidate retrieval.
37
00:04:34,994 --> 00:04:46,254
The main design choice is to expose large 3D scenes through interactively queried visual observations that a vision-language model, or VLM, can inspect and act upon.
38
00:04:46,298 --> 00:04:49,778
Evan: What does this interactive environment look like?
39
00:04:49,826 --> 00:04:57,646
Ashley: The environment represents the scene using a bird’s-eye-view image, or BEV, as the initial global observation.
40
00:04:57,646 --> 00:05:05,396
The VLM agent can interact with the scene through three main tools: Landmark Query, Zoom-In, and Oblique Rendering.
41
00:05:05,396 --> 00:05:11,186
Each tool serves a specific purpose in refining the search region and verifying details.
42
00:05:11,294 --> 00:05:14,044
Evan: Let's break down each of these tools.
43
00:05:14,044 --> 00:05:16,274
What does the Landmark Query do?
44
00:05:16,322 --> 00:05:27,942
Ashley: The Landmark Query resolves a named landmark from the query, such as a building or road name, and returns its world position and its projection in the selected view.
45
00:05:27,942 --> 00:05:37,302
This tool is particularly useful for leveraging geographic features described in referring expressions to establish coarse spatial anchors.
46
00:05:37,416 --> 00:05:40,966
Evan: And how does the Zoom-In tool work?
47
00:05:41,018 --> 00:05:50,328
Ashley: The Zoom-In tool allows the agent to select a region of interest within the BEV view and obtain a high-resolution local crop of that area.
48
00:05:50,328 --> 00:06:02,198
This crop reveals detail that is absent from the downsampled global observation, crucial for distinguishing small and densely packed urban objects, for instance, cars in a parking lot.
49
00:06:02,258 --> 00:06:05,838
Evan: How about the Oblique Rendering tool?
50
00:06:05,882 --> 00:06:20,532
Ashley: The Oblique Rendering tool renders a selected candidate region from a specified cardinal direction, creating a 3D-aware view that helps verify the object's shape, height, and spatial placement.
51
00:06:20,532 --> 00:06:30,222
This tool provides complementary visual evidence to the top-down view, which can be essential for confirming 3D structures and relative positions.
52
00:06:30,326 --> 00:06:35,586
Evan: So, the agent receives the current observation and interaction history.
53
00:06:35,586 --> 00:06:38,586
How does it decide what action to take next?
54
00:06:38,642 --> 00:06:57,592
Ashley: The agent follows a policy, denoted as πθ, which is a VLM that processes the current observation and the interaction history, including previous tool calls and visual observations.
55
00:06:57,592 --> 00:07:10,002
For each step, the policy generates a reasoning trace and an action, which can be either a tool call with structured arguments or a final answer.
56
00:07:10,118 --> 00:07:14,458
Evan: It sounds like a complex decision-making process.
57
00:07:14,458 --> 00:07:17,598
How is it optimized using reinforcement learning?
58
00:07:17,782 --> 00:07:26,322
Ashley: Urbex is trained using reinforcement learning, modeling the grounding episode as a multi-step Markov decision process.
59
00:07:26,322 --> 00:07:35,892
The terminal reward encourages high overlap with the target and provides a dense center-distance signal when overlap is initially sparse.
60
00:07:35,892 --> 00:07:47,442
In addition, tool-level shaping rewards guide intermediate behaviors, such as zooming toward the target region or rendering oblique views that align well with the target.
61
00:07:47,498 --> 00:07:51,188
Evan: But reinforcement learning can be computationally intensive.
62
00:07:51,188 --> 00:07:55,038
How does Urbex ensure efficiency during training and inference?
63
00:07:55,082 --> 00:08:10,592
Ashley: To manage efficiency, Urbex uses Group Relative Policy Optimization, or GRPO, which circumvents the memory overhead of a separate value network by estimating the baseline directly from multiple concurrent rollouts.
64
00:08:10,592 --> 00:08:21,142
The policy is optimized by maximizing a clipped objective function, relying on this reinforcement learning framework without supervised trajectory demonstrations.
65
00:08:21,254 --> 00:08:32,814
Evan: So the environment, tools, and optimization all work together to enhance the agent's ability to search actively and accurately within large city-scale scenes.
66
00:08:32,858 --> 00:08:34,178
Ashley: Exactly.
67
00:08:34,178 --> 00:08:51,178
Urbex integrates these components into a cohesive framework that allows a VLM to dynamically interact with urban environments, leveraging landmarks, refining regions of interest, and verifying details through 3D renderings, all optimized by reinforcement learning.
68
00:08:51,242 --> 00:09:02,682
Evan: How does this method compare with other approaches, especially in terms of performance metrics?
69
00:09:02,808 --> 00:09:11,728
Ashley: Urbex achieves the best Hit@1 and IoU on the CityRefer dataset, outperforming traditional ranking-based methods.
70
00:09:11,728 --> 00:09:18,148
It supports candidate-free grounding and significantly improves online inference efficiency.
71
00:09:18,148 --> 00:09:29,018
The model's ability to achieve strong zero-shot robustness on STPLS3D-Refer further underscores its versatile application.
72
00:09:29,066 --> 00:09:35,906
Evan: It sounds like a comprehensive approach addressing multiple challenges in 3D city-scale grounding.
73
00:09:35,906 --> 00:09:39,046
Is there anything else noteworthy about the evaluation?
74
00:09:39,098 --> 00:09:40,218
Ashley: Indeed.
75
00:09:40,218 --> 00:09:50,968
The evaluation also highlights the scalability and feasibility of Urbex in real-world applications due to its efficient tool use and robust learning policy.
76
00:09:50,968 --> 00:10:00,578
Detailed comparisons, ablation studies, and qualitative examples further demonstrate the model's strengths and areas for potential improvement.
77
00:10:00,626 --> 00:10:02,826
Evan: That wraps up the Method section.
78
00:10:02,826 --> 00:10:09,326
Next, we'll dive into the experiments and results to see how Urbex truly performs.
79
00:10:10,587 --> 00:10:16,297
Evan: Now, let's get to the exciting part: the experiments and results.
80
00:10:16,297 --> 00:10:20,587
Ashley, tell us how the authors evaluated Urbex.
81
00:10:20,683 --> 00:10:25,213
Ashley: Evan, the authors conducted extensive experiments to evaluate Urbex.
82
00:10:25,213 --> 00:10:32,443
They focused on three key areas: in-domain performance, cross-scene generalization, and component sensitivity.
83
00:10:32,499 --> 00:10:36,639
Evan: Which datasets did they use for these evaluations?
84
00:10:36,699 --> 00:10:44,379
Ashley: The primary dataset used was CityRefer,
which is built on SensatUrban, a large-scale urban point cloud dataset.
85
00:10:44,379 --> 00:10:53,919
Additionally, they created a cross-domain benchmark called STPLS3D-Refer to test zero-shot generalization.
86
00:10:53,919 --> 00:10:59,519
This benchmark was derived from the STPLS3D photogrammetry point clouds.
87
00:10:59,571 --> 00:11:05,071
Evan: How did Urbex perform on the CityRefer dataset?
88
00:11:05,115 --> 00:11:13,275
Ashley: On CityRefer, Urbex achieved the best Hit@1 and Intersection over Union (IoU) metrics compared to the baselines.
89
00:11:13,275 --> 00:11:30,695
Specifically, Urbex attained a Hit@1 score of 25.8%, substantially outperforming the strongest city-scale ranking baseline, CityAnchor, which had a Hit@1 score of 23.4% in the ND split and 14.2% in the NO split.
90
00:11:30,747 --> 00:11:34,907
Evan: What about the comparison with other visual grounding methods?
91
00:11:35,091 --> 00:11:37,901
Ashley: Urbex was compared with three groups of methods.
92
00:11:37,901 --> 00:11:48,581
First, general-purpose multimodal models like GPT-5.4 and Qwen3.6-plus, which performed significantly lower in Hit@1 scores.
93
00:11:48,581 --> 00:11:57,781
Second, indoor 3D grounding methods like InstanceRefer and EDA, which struggled significantly due to the complexity and scale of city scenes.
94
00:11:57,781 --> 00:12:07,351
Third, other city-scale grounding methods like CityRefer and CityAnchor, which rely heavily on precomputed candidates and showed lower efficiency and accuracy.
95
00:12:07,455 --> 00:12:11,015
Evan: Efficiency is crucial for practical applications.
96
00:12:11,015 --> 00:12:14,795
How did Urbex perform in terms of inference efficiency?
97
00:12:14,929 --> 00:12:17,319
Ashley: Urbex was substantially more efficient.
98
00:12:17,319 --> 00:12:28,479
While CityAnchor required about 32.1 seconds per example due to scoring numerous candidate instances, Urbex only took about 3.2 seconds per example on average.
99
00:12:28,479 --> 00:12:37,439
This efficiency is largely due to Urbex's method of focusing on a small number of informative regions rather than exhaustively scoring all instances.
100
00:12:37,551 --> 00:12:45,891
Evan: And how did Urbex perform on the zero-shot tests with STPLS3D-Refer?
101
00:12:45,939 --> 00:12:50,679
Ashley: In the zero-shot context, the results were also impressive.
102
00:12:50,679 --> 00:13:08,059
On STPLS3D-Refer, Urbex achieved a Hit@1 score of 27.2%, significantly outperforming GPT-5.4, which had a Hit@1 score of 16.3%.
103
00:13:08,059 --> 00:13:13,739
This indicates that Urbex's active search strategy generalizes well to new scenes.
104
00:13:13,803 --> 00:13:16,703
Evan: Let's talk about their ablation studies.
105
00:13:16,703 --> 00:13:18,603
What insights did they offer?
106
00:13:18,651 --> 00:13:23,071
Ashley: The ablation studies highlighted the importance of each component of Urbex.
107
00:13:23,071 --> 00:13:38,031
For instance, adding Landmark Query significantly boosted Hit@3 from 15.5% to 28.2% and 2D IoU from 4.1% to 8.9%.
108
00:13:38,031 --> 00:13:50,491
Including the Oblique Rendering tool and extending the interaction budget from 7 to 8 turns also improved performance, suggesting that each tool plays a crucial role in the model's success.
109
00:13:50,547 --> 00:13:54,887
Evan: How about the reward design in the reinforcement learning process?
110
00:13:54,939 --> 00:13:58,359
Ashley: The reward design was crucial for effective learning.
111
00:13:58,359 --> 00:14:09,509
Using a combination of terminal rewards for final localization accuracy and intermediate rewards for tool usage and navigation efficiency yielded the best results.
112
00:14:09,509 --> 00:14:26,539
For example, including distance shaping rewards improved Hit@3 from 22.8% to 30.6% and 2D IoU from 6.7% to 10.9%, demonstrating the importance of these guiding signals.
113
00:14:26,655 --> 00:14:31,615
Evan: Did they explore how the model handles more practical deployment scenarios?
114
00:14:31,659 --> 00:14:40,669
Ashley: Yes, they tested Urbex in practical settings by measuring inference time and testing candidate recall with predicted instance proposals.
115
00:14:40,669 --> 00:14:51,419
Their results showed that Urbex maintained high localization quality and efficiently handled practical deployment challenges, making it suitable for real-world applications.
116
00:14:51,483 --> 00:14:59,253
Evan: It sounds like the experiments provide strong evidence supporting Urbex's effectiveness and robustness.
117
00:14:59,253 --> 00:15:01,203
Shall we wrap up this section?
118
00:15:01,251 --> 00:15:02,531
Ashley: Yes, Evan.
119
00:15:02,531 --> 00:15:05,141
That's the end of the Experiment section.
120
00:15:05,141 --> 00:15:15,211
Urbex demonstrated impressive performance across multiple benchmarks and scenarios, showcasing its potential for dynamic and accurate city-scale 3D grounding.
121
00:15:16,529 --> 00:15:27,669
Evan: With the experiments and results clearly showcasing Urbex's effectiveness, let's shift gears to understand how this work fits into the broader research landscape.
122
00:15:27,669 --> 00:15:31,649
Ashley, can you guide us through the related work in this domain?
123
00:15:31,709 --> 00:15:33,029
Ashley: Certainly, Evan.
124
00:15:33,029 --> 00:15:42,929
The Related Work section places Urbex within the context of existing methodologies and the broader field of 3D scene grounding and vision-language models.
125
00:15:42,989 --> 00:15:46,339
Evan: Let's start with 3D scene grounding.
126
00:15:46,339 --> 00:15:50,729
What have previous methods focused on, and how does Urbex differ?
127
00:15:50,789 --> 00:15:56,489
Ashley: Most prior work on 3D scene grounding has centered around indoor environments.
128
00:15:56,489 --> 00:16:08,839
These methods typically involve encoding point clouds or multi-view images, merging them with language encoders, and ranking candidate instances via cross-modal matching.
129
00:16:08,839 -->
00:16:23,629
Notable advances include integrating relation modeling, multi-view context, and transformer-based architectures, as well as exploring bottom-up localization approaches to avoid errors from segmentation models.
130
00:16:23,753 --> 00:16:29,313
Evan: And how does this translate to the challenges of city-scale scenes?
131
00:16:29,357 --> 00:16:39,737
Ashley: When applied to city-scale grounding, outdoor urban scenes pose unique challenges due to their vast spatial extent and repetitive elements like vehicles and buildings.
132
00:16:39,737 --> 00:16:48,217
Existing city-scale methods often adopt a static candidate-ranking pipeline, scoring precomputed instances during inference.
133
00:16:48,217 --> 00:16:57,927
While this avoids processing the entire point cloud, it is heavily reliant on the quality of these proposals and can become inefficient with numerous candidates.
134
00:16:57,927 --> 00:17:05,017
Urbex, in contrast, reframes this problem as an active spatial search, dynamically interacting with the scene.
135
00:17:05,069 --> 00:17:06,959
Evan: That makes a lot of sense.
136
00:17:06,959 --> 00:17:10,389
How about the role of vision-language models in this context?
137
00:17:10,445 --> 00:17:20,915
Ashley: Vision-language models, or VLMs, have made remarkable strides in tasks involving visual question answering, fine-grained reasoning, and instruction following.
138
00:17:20,915 --> 00:17:25,325
Reinforcement learning has further enhanced these models' reasoning capabilities.
139
00:17:25,325 --> 00:17:36,525
The recent trend has shifted from passive perception to more agentic systems capable of using tools, manipulating interfaces, and navigating complex visual environments.
140
00:17:36,525 --> 00:17:49,245
Urbex builds on this progression, leveraging these advanced VLMs to actively interact with 3D city scenes and refine their understanding through tools like Landmark Query, Zoom-In, and Oblique Rendering.
141
00:17:49,301 --> 00:17:59,821
Evan: Are there similar agentic frameworks that Urbex draws inspiration from or builds upon?
142
00:17:59,885 --> 00:18:01,495
Ashley: Yes, indeed.
143
00:18:01,495 --> 00:18:09,335
Urbex aligns with recent research transitioning from 2D tasks and indoor navigation to more complex 3D settings.
144
00:18:09,335 --> 00:18:18,445
Similar works include dynamic 2D-to-3D reconstruction using visual tools and active cognition over open-world environments.
145
00:18:18,445 --> 00:18:27,315
However, these approaches primarily operate in indoor or graph-structured settings and are not specifically designed for city-scale environments.
146
00:18:27,315 --> 00:18:38,365
Urbex specializes this agentic paradigm by providing a suite of city-aware spatial tools and a reinforcement learning objective for efficient exploration in vast urban areas.
147
00:18:38,429 --> 00:18:44,089
Evan: How does Urbex's evaluation strategy compare to those used in previous works?
148
00:18:44,141 --> 00:18:57,611
Ashley: Evaluation in prior works has often focused on zero-shot and open-world methodologies, utilizing foundation models, multi-view reasoning, and scene graphs to reduce dependency on task-specific supervision.
149
00:18:57,611 --> 00:19:12,001
Urbex follows a similar spirit but advances it by testing on both in-domain and cross-domain datasets—namely CityRefer for primary evaluation and STPLS3D-Refer for zero-shot transfer testing.
150
00:19:12,001 --> 00:19:20,201
This rigorous approach ensures that Urbex is not only effective within the training domain but also robust across varied urban scenes.
151
00:19:20,261 --> 00:19:28,111
Evan: It sounds like Urbex stands on the shoulders of many notable advances in both 3D grounding and vision-language interaction.
152
00:19:28,111 --> 00:19:32,121
Any final thoughts on how this body of work influences Urbex?
153
00:19:32,165 --> 00:19:33,655
Ashley: Exactly, Evan.
154
00:19:33,655 --> 00:19:42,155
Urbex leverages and extends prior innovations by integrating dynamic spatial interaction and tool use within large-scale urban environments.
155
00:19:42,155 --> 00:19:50,645
This synthesis of past advances and new contributions positions Urbex as a significant step forward in the domain of 3D scene grounding.
156
00:19:50,753 --> 00:19:51,583
Evan: Great.
157
00:19:51,583 --> 00:19:54,823
That wraps up our discussion on the Related Work section.
158
00:19:54,823 --> 00:20:00,393
Up next, we'll dive into the methods Urbex uses to achieve its impressive results.
159
00:20:01,638 --> 00:20:05,878
Evan: We've covered the methods, experiments, and related work.
160
00:20:05,878 --> 00:20:11,278
Ashley, let's summarize the key contributions and takeaways of Urbex.
161
00:20:11,374 --> 00:20:12,244
Ashley: Evan.
162
00:20:12,244 --> 00:20:20,164
Urbex represents a significant leap in 3D city-scale grounding by shifting from static candidate ranking to active spatial search.
163
00:20:20,164 --> 00:20:30,314
The framework leverages a vision-language model to interact dynamically with large urban environments, using tools like Landmark Query, Zoom-In, and Oblique Rendering.
164
00:20:30,426 --> 00:20:36,126
Evan: It's clear that this dynamic approach allows for more precise and efficient grounding.
165
00:20:36,174 --> 00:20:37,034
Ashley: Indeed.
166
00:20:37,034 --> 00:20:49,804
Urbex's key contributions include the innovative formulation of grounding as an active spatial search, the design of city-specific interactive tools, and the optimization of the agent's search policy through reinforcement learning.
167
00:20:49,804 --> 00:21:02,754
The model achieves impressive performance metrics, such as the highest Hit@1 and Intersection over Union scores on the CityRefer dataset and strong zero-shot robustness on STPLS3D-Refer.
168
00:21:02,814 --> 00:21:14,054
Evan: This method not only improves accuracy and efficiency but also offers practical scalability for real-world applications like autonomous navigation and urban planning.
169
00:21:14,118 --> 00:21:26,438
Ashley: Exactly, Urbex efficiently handles the complexities of large-scale urban scenes, supports candidate-free grounding, and showcases impressive cross-scene generalization.
170
00:21:26,438 --> 00:21:34,678
Its reinforcement learning framework further enhances the model's ability to navigate and interact with these environments dynamically.
171
00:21:34,734 --> 00:21:38,224
Evan: That brings us to the end of today's episode.
172
00:21:38,224 --> 00:21:41,194
Ashley, any final notes for our listeners?
173
00:21:41,268 --> 00:21:48,278
Ashley: I just want to highlight the importance of continuing to innovate and explore dynamic approaches in AI research.
174
00:21:48,278 --> 00:21:54,198
Urbex is a great example of how new methodologies can push the boundaries of what we can achieve.
175
00:21:54,316 --> 00:21:55,566
Evan: Thank you, Ashley.
176
00:21:55,566 --> 00:21:58,266
And thank you to our listeners for joining us today.
177
00:21:58,266 --> 00:22:00,766
We hope you found this discussion insightful.
178
00:22:00,822 --> 00:22:06,752
Ashley: Be sure to tune in for our next episode, where we'll dive into more exciting AI research.
179
00:22:06,752 --> 00:22:11,622
Until then, stay curious and keep exploring the world of AI!
180
00:22:11,710 --> 00:22:13,930
Evan: Goodbye for now, everyone!