1
00:00:03,060 --> 00:00:05,620
Evan: Welcome to Daily Paper Cast.

2
00:00:05,724 --> 00:00:14,804
Evan: Today's paper is selected from the Hugging Face daily paper list for October 13, 2026, and it has garnered 107 upvotes.

3
00:00:14,856 --> 00:00:23,976
Ashley: The title of the paper we're discussing today is 'GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives.'

4
00:00:24,024 --> 00:00:31,624
Evan: The first two authors of this paper are Qize Yu and Lianrui Fan, with Ping Luo as the corresponding author.

5
00:00:31,624 --> 00:00:37,884
They are affiliated with XPeng Inc., Peking University, and The University of Hong Kong, respectively.

6
00:00:37,944 --> 00:00:41,064
Ashley: Let's dive into the Introduction section of the paper.

7
00:00:41,112 --> 00:00:55,112
Evan: Visual grounding, the primary focus of this paper, connects language to objects, locations, and structures that are relevant for interaction, making it a core perceptual capability of vision-language models or VLMs.

8
00:00:55,176 --> 00:01:06,476
Ashley: Despite recent improvements in grounding user interface elements, regions, and task-relevant entities, achieving broad and precise grounding for physical intelligence remains challenging.

9
00:01:06,528 --> 00:01:17,028
Evan: To illustrate this, consider the task of organizing a cluttered desk: one needs to locate the mug and its handle amidst the clutter before deciding how to grasp and move it.

10
00:01:17,088 --> 00:01:22,128
Ashley: Current vision-language-action models and world-action models have their limitations.

11
00:01:22,128 --> 00:01:31,918
These models often use general-purpose vision-language and video-generation backbones, which aren't optimized for the precise grounding required by physical interaction.

12
00:01:31,918 --> 00:01:36,888
This inadequacy in perception can become a bottleneck for downstream action learning.

13
00:01:36,996 --> 00:01:51,216
Evan: Indeed, evaluations of general-purpose VLMs reveal weaknesses in demanding settings, such as dense scenes and tiny objects, where even leading models like Qwen3.5 and GPT-6 Astra struggle.

14
00:01:51,264 --> 00:01:58,264
Ashley: These perceptual gaps mean that limited action data must simultaneously handle perceptual and action learning.

15
00:01:58,264 --> 00:02:05,024
Additionally, compute and latency constraints further limit reliance on larger backbones for action execution.

16
00:02:05,088 --> 00:02:11,508
Evan: This has led the authors to propose building a more capable grounding foundation model to alleviate these issues.

17
00:02:11,568 --> 00:02:18,788
Ashley: The paper presents three key questions: How far can we push the perceptual capabilities of grounding foundation models?

18
00:02:18,788 --> 00:02:25,558
What does a stronger grounding foundation contribute to physical intelligence, and which grounding capabilities matter most?

19
00:02:25,558 --> 00:02:31,768
And finally, what might these findings suggest about the design of future embodied foundation models?

20
00:02:31,824 --> 00:02:39,334
Evan: GroundingPI, a proposed 4-billion-parameter grounding foundation model, aims to address these questions.

21
00:02:39,334 --> 00:02:44,364
This model utilizes visual primitives including points and bounding boxes.

22
00:02:44,454 --> 00:02:53,394
Ashley: GroundingPI unifies diverse perception tasks as language-conditioned structured generation through a shared vocabulary with quantized coordinates.

23
00:02:53,394 --> 00:03:04,944
The model covers grounding, referring, pointing, optical character recognition or OCR, graphical user interface or GUI, layout grounding, and visual prompting.

24
00:03:04,992 --> 00:03:19,672
Evan: Through multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with group relative policy optimization, GroundingPI establishes a new state of the art across 34 benchmarks and 44 baselines.

25
00:03:19,728 --> 00:03:29,678
Ashley: It averages 73.68% performance, exceeding the larger GPT-6 Astra model, which averages 71.54%.

26
00:03:29,678 --> 00:03:36,168
GroundingPI improves performance notably in robotic manipulation and autonomous driving tasks.

27
00:03:36,216 --> 00:03:47,756
Evan: For instance, on RoboTwin 2.0, it outperforms mainstream backbones by up to 24.8% relative to the strongest baseline across four out-of-distribution settings.

28
00:03:47,756 --> 00:03:57,636
On RoboCasa-GR1, GroundingPI trained with 50% of the demonstrations outperforms baselines trained with 75% of the demonstrations.

29
00:03:57,696 --> 00:04:10,676
Ashley: The paper systematically analyzes GroundingPI's pretraining in terms of scale and data composition, revealing substantial benefits for dense grounding and OCR's potential as a catalyst for perceptual learning.

30
00:04:10,728 --> 00:04:22,928
Evan: These findings ultimately support the idea of grounding as a perceptual foundation and dedicated perceptual pretraining as a promising direction for foundation models of physical intelligence.

31
00:04:22,992 --> 00:04:26,062
Ashley: That brings us to the end of the Introduction section.

32
00:04:26,062 --> 00:04:29,632
Let's now move on to the methodology proposed in this paper.

33
00:04:30,890 --> 00:04:35,610
Evan: Let's delve into the methodology used in this paper to develop GroundingPI.

34
00:04:35,666 --> 00:04:37,026
Ashley: Sure, Evan.

35
00:04:37,026 --> 00:04:43,326
GroundingPI is an autoregressive grounding foundation model designed for language-conditioned perception.

36
00:04:43,326 --> 00:04:47,766
The model architecture integrates multiple components to achieve its goals.

37
00:04:47,810 --> 00:04:58,970
Evan: Interestingly, the model combines a visual encoder named MoonViT-V2, a learnable multimodal projector, and a Qwen3-4B language backbone.

38
00:04:58,970 --> 00:05:01,450
Could you explain how these components interact?

39
00:05:01,514 --> 00:05:09,524
Ashley: As shown in the paper, the visual encoder processes the image and produces visual features, which are then aggregated by the projector.

40
00:05:09,524 --> 00:05:20,434
The aggregated visual features are mapped into the language embedding space through a two-layer Multilayer Perceptron or MLP with GELU activation and output normalization.

41
00:05:20,498 --> 00:05:30,988
Evan: So the visual encoder handles the raw image processing, the projector prepares these features for interaction with the language model, and the language backbone generates the final responses.

42
00:05:30,988 --> 00:05:35,898
How does the model ensure the coordinated generation of visual and linguistic tokens?

43
00:05:35,954 --> 00:05:45,314
Ashley: The model uses an autoregressive decoder to generate a variable-length response, conditioned on both visual embeddings from the encoder and the input query.

44
00:05:45,314 --> 00:05:55,274
This response includes semantic labels, protocol markers, and quantized coordinate tokens, integrating all aspects of the grounded perception tasks seamlessly.

45
00:05:55,322 --> 00:06:00,952
Evan: That makes sense, but managing multiple perception tasks must involve different kinds of data.

46
00:06:00,952 --> 00:06:03,742
How does GroundingPI handle its training data?

47
00:06:03,824 --> 00:06:10,254
Ashley: GroundingPI's training data combines public datasets with annotations produced by a sophisticated data engine.

48
00:06:10,254 --> 00:06:12,214
It's quite fascinating, actually.

49
00:06:12,214 --> 00:06:21,054
The data engine uses fusion from multiple teacher annotations at the field level and applies task-specific validation to ensure high-quality labels.

50
00:06:21,098 --> 00:06:26,198
Evan: And the validated labels are used iteratively to improve the model's precision, right?

51
00:06:26,258 --> 00:06:27,268
Ashley: Exactly.

52
00:06:27,268 --> 00:06:36,718
Accepted labels train a unified grounding expert in iterative cycles, with complementary teachers and local observations resolving any uncertain cases.

53
00:06:36,718 --> 00:06:41,718
This process refines supervision and continuously improves the labeling accuracy.

54
00:06:41,822 --> 00:06:43,832
Evan: Let's talk about the training design.

55
00:06:43,832 --> 00:06:51,732
The paper outlines three main stages: base VLM training, supervised fine-tuning, and reinforcement post-training.

56
00:06:51,732 --> 00:06:53,902
Could you break these stages down for us?

57
00:06:53,954 --> 00:06:54,944
Ashley: Certainly.

58
00:06:54,944 --> 00:07:02,784
Base training begins with aligning the projector using image-caption supervision while freezing the visual encoder and language backbone.

59
00:07:02,784 --> 00:07:10,714
This stage focuses on aligning visual features with the language input space without any feature-distance or contrastive objectives.

60
00:07:10,778 --> 00:07:19,158
Evan: Following this alignment, the model undergoes joint multimodal pretraining, which involves text-only and general image-text data.

61
00:07:19,158 --> 00:07:21,038
What's unique about this stage?

62
00:07:21,098 --> 00:07:31,158
Ashley: During this stage, all model components are updated jointly, leveraging text examples for causal supervision and multimodal examples for assistant responses.

63
00:07:31,158 --> 00:07:36,518
This helps in fine-tuning the interactions between the visual and language sides of the model.

64
00:07:36,578 --> 00:07:40,818
Evan: The third stage involves general visual and video understanding.

65
00:07:40,818 --> 00:07:42,138
What does this entail?

66
00:07:42,194 --> 00:07:48,334
Ashley: In this stage, the model receives supervision from visual question answering and instruction-following data.

67
00:07:48,334 --> 00:07:55,854
It also processes image captions and videos, extending the context and packing lengths to accommodate longer multimodal sequences.

68
00:07:55,854 --> 00:08:00,774
This enhancement is crucial for understanding complex interactions and scenarios.

69
00:08:00,858 --> 00:08:07,338
Evan: After these three stages comes supervised fine-tuning, sometimes referred to as Pretrain 2.

70
00:08:07,338 --> 00:08:08,698
What's the focus here?

71
00:08:08,762 --> 00:08:13,392
Ashley: Supervised fine-tuning is where the model specializes in spatial perception.

72
00:08:13,392 --> 00:08:18,292
It focuses on coordinating visual location alignment and shared output protocols.

73
00:08:18,292 --> 00:08:26,382
Under teacher forcing, it minimizes token-level cross-entropy across semantic labels, coordinate tokens, and protocol markers.

74
00:08:26,426 --> 00:08:33,686
Evan: This fine-tuning stage ensures the model can accurately generate structured outputs even under varying conditions.

75
00:08:33,686 --> 00:08:36,006
What about reinforcement post-training?

76
00:08:36,050 --> 00:08:45,470
Ashley: Reinforcement post-training employs Group Relative Policy Optimization, or GRPO, to directly optimize grounding completeness and localization.

77
00:08:45,470 --> 00:08:57,690
By sampling multiple responses per image-query pair, it measures coverage and localization accuracy for grounding tasks, while for OCR tasks, it evaluates text-to-geometry consistency.

78
00:08:57,746 --> 00:08:59,766
Evan: That's a comprehensive approach.

79
00:08:59,766 --> 00:09:07,846
The rewards in GRPO are standardized within the prompt and combined to update the model parameters through clipping and optimization.

80
00:09:07,846 --> 00:09:13,086
How does this reinforcement training contribute to the model's performance in downstream tasks?

81
00:09:13,130 --> 00:09:29,230
Ashley: Through reinforcement post-training, the model optimizes both grounding accuracy and text geometry consistency, which directly enhances its ability to perform precise and reliable physical intelligence tasks like robotic manipulation and autonomous driving.

82
00:09:29,282 --> 00:09:36,002
Evan: It's clear that GroundingPI's training involves iterative improvement and sophisticated optimization techniques.

83
00:09:36,002 --> 00:09:43,142
The model seems to be designed with robust methodologies that make it particularly capable in diverse grounding scenarios.

84
00:09:43,202 --> 00:09:44,372
Ashley: Indeed, Evan.

85
00:09:44,372 --> 00:09:57,432
GroundingPI leverages a mix of multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning, making it a powerful tool in advancing perceptual capabilities in physical intelligence.

86
00:09:57,432 --> 00:10:01,122
And that concludes our discussion on the Method section of the paper.

87
00:10:02,379 --> 00:10:10,819
Evan: Next, let's move on to the Experiment and Results section, where the authors evaluate GroundingPI across a range of benchmarks.

88
00:10:10,875 --> 00:10:12,245
Ashley: Indeed, Evan.

89
00:10:12,245 --> 00:10:21,245
The authors set out to answer three core questions by conducting several experiments: How far can GroundingPI advance broad and precise perception?

90
00:10:21,245 --> 00:10:24,875
Does stronger perception support stronger physical intelligence?

91
00:10:24,875 --> 00:10:31,375
Which perceptual capabilities help explain this transfer, and what might they suggest for embodied foundations?

92
00:10:31,419 --> 00:10:40,929
Evan: To address these questions, GroundingPI is evaluated against 44 baselines across 34 benchmarks covering 11 perceptual capabilities.

93
00:10:40,929 --> 00:10:48,879
This includes specialized detectors, general-purpose vision-language models, grounding specialists, and embodied foundations.

94
00:10:48,939 --> 00:11:16,739
Ashley: Yes, and the evaluation suite spans a broad array of tasks such as common and long-tailed detection on COCO and LVIS datasets, dense and tiny object detection on Dense200 and VisDrone, referring grounding on RefCOCO, RefCOCO+, RefCOCOg, and HumanRef, and spatial and GUI grounding using RoboSpatial, RefSpatial, ScreenSpot-V2, ScreenSpot-Pro, and OSWorld-G.

95
00:11:16,863 --> 00:11:37,203
Evan: It seems like they also evaluated OCR tasks using HierText, ICDAR2015, TotalText, and SROIE, layout grounding with DocLayNet and M6Doc, and exemplar-based visual prompting on FSC147 and the detection benchmarks mentioned earlier.

96
00:11:37,251 --> 00:11:38,421
Ashley: Precisely.

97
00:11:38,421 --> 00:11:53,921
Main results indicate that GroundingPI achieves a 73.68% average performance, outperforming similarly sized models and remaining competitive with GPT-6 Astra, which averages 71.54%.

98
00:11:53,921 --> 00:11:58,871
This is a strong indication of its robust capabilities across various tasks.

99
00:11:58,923 --> 00:12:11,673
Evan: When looking into specific benchmarks like Dense200, GroundingPI reaches 74.53 in F1mIoU, which directly addresses the dense-scene perceptual gap.

100
00:12:11,673 --> 00:12:24,563
In contrast, Astra achieves 65.04 and Qwen3-VL-4B only 14.02, highlighting GroundingPI’s superior handling of densely packed objects.

101
00:12:24,627 --> 00:12:31,177
Ashley: Further, in the robot and spatial grounding tasks, GroundingPI shows strong spatial transfer capabilities.

102
00:12:31,177 --> 00:12:36,347
On RefSpatial benchmarks, it averages high accuracy on Location and Placement splits.

103
00:12:36,347 --> 00:12:43,527
GroundingPI's performance on these splits indicates reliable spatial understanding, which is crucial for robotic interaction.

104
00:12:43,587 --> 00:12:44,747
Evan: That's impressive.

105
00:12:44,747 --> 00:12:47,627
And what about OCR and document layout tasks?

106
00:12:47,627 --> 00:12:53,067
Given the complexity of text recognition and layout analysis, how does GroundingPI fare?

107
00:12:53,185 --> 00:13:10,935
Ashley: OCR and document layout tasks are indeed complex, but GroundingPI performs exceptionally well, attaining 72.47 on SROIE and 74.82 on M6Doc, compared to Astra’s 53.57 and 60.59.

108
00:13:10,935 --> 00:13:18,375
These results extend to precise localization of document regions and accurate text recognition across diverse formats.

109
00:13:18,465 --> 00:13:24,235
Evan: GroundingPI’s capabilities seem well-rounded, but what about tasks involving object pointing?

110
00:13:24,291 --> 00:13:29,721
Ashley: For object pointing, GroundingPI leads in six of the seven benchmark columns.

111
00:13:29,721 --> 00:13:41,371
Its precision and recall metrics are notably higher compared to leading models like Astra and Rex-Omni, suggesting reliable instance identification and localization within the predicted masks.

112
00:13:41,427 --> 00:13:42,817
Evan: They didn’t stop there.

113
00:13:42,817 --> 00:13:52,317
Their experiments also probed GroundingPI’s transferability to physical intelligence tasks like autonomous driving and robotic manipulation.

114
00:13:52,317 --> 00:13:54,287
What did they find in these areas?

115
00:13:54,339 --> 00:14:08,319
Ashley: In autonomous driving tasks evaluated on nuScenes, GroundingPI achieves an average open-loop L2 error of 0.296 meters, improving on Qwen3-VL-4B and RynnBrain baselines.

116
00:14:08,319 --> 00:14:23,679
For robotic manipulation tasks evaluated on RoboTwin 2.0 and RoboCasa-GR1, GroundingPI ranks favorably in five of six manipulation settings, highlighting its strong performance in both ID and OOD settings.

117
00:14:23,739 --> 00:14:26,329
Evan: How does data efficiency play into this?

118
00:14:26,329 --> 00:14:30,679
Are there notable findings regarding action-learning from limited demonstrations?

119
00:14:30,723 --> 00:14:31,793
Ashley: Indeed.

120
00:14:31,793 --> 00:14:38,963
GroundingPI's data-efficiency evaluations demonstrate its capacity for learning efficiently from limited demonstrations.

121
00:14:38,963 --> 00:14:54,123
For example, using just 50% of the RoboCasa-GR1 demonstrations, GroundingPI surpasses all baseline models trained with 75% of the demonstrations, showcasing its efficient action learning capability from scarce data.

122
00:14:54,171 --> 00:14:56,701
Evan: That’s an extraordinary efficiency.

123
00:14:56,701 --> 00:15:01,321
Scaling and data composition also seem to be crucial for model performance.

124
00:15:01,321 --> 00:15:04,771
What conclusions did the authors draw regarding these aspects?

125
00:15:04,827 --> 00:15:11,717
Ashley: The authors found that scaling grounding pretraining enhances both perceptual quality and downstream action learning.

126
00:15:11,717 --> 00:15:23,447
Additionally, dense grounding proves essential for spatial perception, while OCR can act as a perceptual catalyst, refining the visual representation and aiding in complex task learning.

127
00:15:23,549 --> 00:15:29,419
Evan: All these findings reinforce the notion of grounding as a foundational perceptual capability.

128
00:15:29,419 --> 00:15:37,259
GroundingPI not only addresses perceptual gaps but also supports efficient and effective physical intelligence tasks.

129
00:15:37,353 --> 00:15:38,123
Ashley: Evan.

130
00:15:38,123 --> 00:15:46,543
This holistic approach to pretraining, fine-tuning, and reinforcement learning makes GroundingPI stand out in advancing physical intelligence.

131
00:15:46,647 --> 00:15:50,337
Evan: And that brings us to the end of the Experiment and Results section.

132
00:15:50,337 --> 00:15:57,527
The insights gained from these structured evaluations lay a solid foundation for the promising future of grounding models.

133
00:15:58,781 --> 00:16:09,801
Evan: Let's move on to the Related Work section, where the authors review the existing literature and position their contributions in the broader context of grounding models and physical intelligence.

134
00:16:09,845 --> 00:16:11,185
Ashley: Certainly, Evan.

135
00:16:11,185 --> 00:16:19,665
The section begins by discussing visual grounding models, which have significantly evolved from the early days of closed-set object detection.

136
00:16:19,709 --> 00:16:32,549
Evan: Yes, models like You Only Look Once or YOLO and DEtection TRansformers or DETR were crucial in this evolution, but they focused primarily on closed-set object detection.

137
00:16:32,597 --> 00:16:33,547
Ashley: Exactly.

138
00:16:33,547 --> 00:16:42,417
YOLO, introduced by Redmon and colleagues in 2016, offered real-time object detection by framing the task as a single regression problem.

139
00:16:42,417 --> 00:16:45,177
DETR, introduced by Carion et al.

140
00:16:45,177 --> 00:16:54,417
in 2020, brought transformers into the mix for end-to-end object detection, but both were primarily oriented towards predefined categories.

141
00:16:54,461 --> 00:16:59,441
Evan: How did the field evolve toward more general and flexible visual grounding models?

142
00:16:59,501 --> 00:17:06,601
Ashley: Progress was marked by the development of language-conditioned localization models, which move beyond fixed vocabularies.

143
00:17:06,601 --> 00:17:20,141
Grounded image-text pretraining, as discussed by Li, Liu, and others in recent works, leverages large-scale data to train models that can understand and localize objects even without predefined categories.

144
00:17:20,189 --> 00:17:26,919
Evan: Generative grounding models seem to adopt an autoregressive approach for label and coordinate prediction.

145
00:17:26,919 --> 00:17:28,549
Could you elaborate on this?

146
00:17:28,613 --> 00:17:38,943
Ashley: Chen, Xiao, and Jiang have explored autoregressive next-token prediction methods to generate structured sequences combining labels and quantized coordinates.

147
00:17:38,943 --> 00:17:46,673
Models such as LocateAnything further advanced precise grounding by incorporating parallel box decoding techniques.

148
00:17:46,733 --> 00:17:48,173
Evan: That sounds innovative.

149
00:17:48,173 --> 00:17:53,463
Additionally, we have seen models extending these capabilities toward robotic interaction.

150
00:17:53,463 --> 00:17:56,613
How does spatial affordance prediction fit into this puzzle?

151
00:17:56,669 --> 00:18:09,319
Ashley: Spatial affordance prediction and spatial referring extend visual grounding to practical applications in robotics, allowing models to infer how objects can be manipulated or interact within a given space.

152
00:18:09,319 --> 00:18:17,249
Works by Yuan, Zhou, and others address these extensions, pushing the boundary of grounding beyond simple object detection.

153
00:18:17,309 --> 00:18:20,969
Evan: Let's pivot to foundation models for physical intelligence.

154
00:18:20,969 --> 00:18:25,319
These models seem to integrate multiple modalities and experiences.

155
00:18:25,319 --> 00:18:27,089
What's the core idea here?

156
00:18:27,149 --> 00:18:33,139
Ashley: Physical intelligence models increasingly incorporate priors obtained from large-scale pretraining.

157
00:18:33,139 --> 00:18:46,449
Vision-Language-Action or VLA models leverage visual-semantic knowledge and language grounding, while World-Action Models or WAMs adapt video and world models to incorporate spatiotemporal dynamics.

158
00:18:46,493 --> 00:18:51,593
Evan: So, integrating semantic, dynamic, and embodied experiences is key.

159
00:18:51,593 --> 00:18:56,573
How do models like those from NVIDIA and Robot Intelligence extend these capabilities?

160
00:18:56,621 --> 00:18:57,761
Ashley: That's right.

161
00:18:57,761 --> 00:19:07,501
Models from these initiatives aim to combine semantic understanding, dynamic context, and embodied experiences to form more general physical intelligence systems.

162
00:19:07,501 --> 00:19:17,401
However, they often face limitations in precise task-conditioned perception needed for physical interaction, leaving gaps that models like GroundingPI aim to fill.

163
00:19:17,453 --> 00:19:21,983
Evan: The paper also discusses perceptual learning and its transfer to action.

164
00:19:21,983 --> 00:19:26,133
Could you highlight some recent policy recipes that integrate these ideas?

165
00:19:26,189 --> 00:19:35,909
Ashley: Recent approaches incorporate bounding-box prediction, region reconstruction, and affordance representations as intermediate steps in action learning.

166
00:19:35,909 --> 00:19:50,529
For example, works by Intelligence et al., Song, Tu, and You utilize spatial auxiliary supervision to strengthen perception, albeit usually as task-specific objectives rather than foundational capabilities.

167
00:19:50,573 --> 00:19:55,273
Evan: What are some limitations of these task-specific perceptual learning strategies?

168
00:19:55,325 --> 00:20:04,635
Ashley: These strategies often require additional annotations, policy-specific components, and do not scale well across diverse scenes and perceptual demands.

169
00:20:04,635 --> 00:20:13,225
This creates challenges in generalizing perception-native foundation models capable of broad and precise perception before action adaptation.

170
00:20:13,277 --> 00:20:16,077
Evan: So, what is the proposed alternative?

171
00:20:16,133 --> 00:20:23,773
Ashley: The authors argue that a more natural and scalable approach is developing strong grounding as a native capability of the foundation model.

172
00:20:23,773 --> 00:20:34,293
By learning broad and precise perception initially, downstream policies can build on solid perceptual foundations rather than cobbling together task-specific proxies.

173
00:20:34,379 --> 00:20:41,469
Evan: GroundingPI fits into this proposal by offering dense grounding and OCR as core pretraining tasks.

174
00:20:41,469 --> 00:20:45,789
This suggests perceptual learning can be foundational rather than ancillary.

175
00:20:45,845 --> 00:20:59,015
Ashley: Indeed, by prioritizing foundational grounding capabilities, models like GroundingPI can better support complex downstream tasks in physical intelligence, including robotic manipulation and autonomous driving.

176
00:20:59,015 --> 00:21:02,825
And with that, we've covered the Related Work section of the paper.

177
00:21:04,146 --> 00:21:09,086
Evan: As we wrap up, let's summarize the key contributions and takeaways from this paper.

178
00:21:09,180 --> 00:21:19,500
Ashley: To start, GroundingPI introduces a robust grounding foundation model that skillfully integrates visual primitives with advanced perceptual tasks.

179
00:21:19,500 --> 00:21:31,290
This innovative approach allows it to excel across a wide range of benchmarks and significantly outperform many leading models, including the larger GPT-6 Astra.

180
00:21:31,390 --> 00:21:43,910
Evan: A major highlight is GroundingPI's ability to lead in both precision and recall across various grounding tasks, including dense and tiny object detection, spatial grounding, and OCR.

181
00:21:43,910 --> 00:21:58,270
Its architectural design, combining the MoonViT-V2 visual encoder, a learnable multimodal projector, and the Qwen3-4B language backbone, ensures a seamless integration of visual and language components.

182
00:21:58,326 --> 00:22:11,456
Ashley: Moreover, the model's training pipeline is meticulously structured into three primary stages: base VLM training, supervised fine-tuning, and reinforcement post-training with Group Relative Policy Optimization.

183
00:22:11,456 --> 00:22:21,426
This comprehensive approach optimizes the model's grounding accuracy and text-geometry consistency, enhancing its capabilities in downstream physical intelligence tasks.

184
00:22:21,486 --> 00:22:30,666
Evan: GroundingPI also excels in data efficiency, particularly in physical intelligence tasks such as robotic manipulation and autonomous driving.

185
00:22:30,666 --> 00:22:37,586
It demonstrates remarkable performance even when trained with reduced data, showcasing its efficient learning capabilities.

186
00:22:37,638 --> 00:22:40,548
Ashley: The implications of these findings are profound.

187
00:22:40,548 --> 00:22:50,888
By prioritizing dense grounding and OCR in the pretraining phase, GroundingPI highlights the potential of perceptual pretraining to support complex downstream tasks.

188
00:22:50,888 --> 00:22:57,038
This positions grounding as a fundamental perceptual capability for future embodied foundation models.

189
00:22:57,102 --> 00:23:06,442
Evan: That concludes our discussion on GroundingPI, a pioneering model pushing the boundaries of physical intelligence through advanced grounding techniques.

190
00:23:06,486 --> 00:23:10,586
Ashley: Thank you for joining us on this episode of Daily Paper Cast.

191
00:23:10,638 --> 00:23:20,618
Evan: We hope you found today's discussion insightful and invite you to join us again for more explorations of cutting-edge research in AI and related fields.

192
00:23:20,720 --> 00:23:27,390
Ashley: Don't forget to check the Hugging Face daily paper list for more fascinating papers and stay tuned for our next episode.

193
00:23:27,478 --> 00:23:31,418
Evan: Until next time, stay curious and keep innovating!

194
00:23:31,500 --> 00:23:32,650
Ashley: Goodbye!