1 00:00:03,000 --> 00:00:05,660 Evan: Welcome to Daily Paper Cast. 2 00:00:05,712 --> 00:00:13,152 Ashley: Today's paper is from the Hugging Face daily paper list of October 2, 2026, with 44 upvotes. 3 00:00:13,200 --> 00:00:23,520 Evan: The paper is titled 'Agent Priors-Guided Policy Learning.' The first two authors are Puming Jiang and Tianrun Hu, and the corresponding author is Harold Soh. 4 00:00:23,520 --> 00:00:26,160 They are from the National University of Singapore. 5 00:00:26,208 --> 00:00:28,278 Ashley: Let's dive into the introduction. 6 00:00:28,278 --> 00:00:37,728 The paper discusses how robots that learn from a few demonstrations often require two types of generalization: compositional and skill generalization. 7 00:00:37,836 --> 00:00:47,476 Evan: Compositional generalization involves recombining skills to solve new tasks, while skill generalization lets the learned policy work in new situations. 8 00:00:47,520 --> 00:00:50,870 Ashley: These types of generalizations are interdependent. 9 00:00:50,870 --> 00:00:56,090 However, information is often lost between the task level and the skill level. 10 00:00:56,090 --> 00:01:05,360 This is because the task level deals with goals and sequences, whereas the skill level deals with specific policies trained from a few demonstrations. 11 00:01:05,484 --> 00:01:10,074 Evan: Existing approaches have tried to connect the task and skill levels in different ways. 12 00:01:10,074 --> 00:01:21,364 For example, vision-language-action models couple task semantics and low-level control in a single model but still face limitations in robustness when layouts and objects shift. 13 00:01:21,438 --> 00:01:27,648 Ashley: Another example is task and motion planning, which connects them through symbolic preconditions and effects. 14 00:01:27,648 --> 00:01:34,068 These are either specified by hand or learned from data, but transitions between skills can still fail. 15 00:01:34,128 --> 00:01:46,128 Evan: Agentic systems can call policies as tools and use language as the interface, yet they often struggle to ground their decisions in the tool's actual ability, leading to failures during transitions. 16 00:01:46,176 --> 00:01:53,296 Ashley: The key idea of this paper is to use each skill's structural prior as part of the interface between composition and the skill. 17 00:01:53,296 --> 00:01:57,476 A structural prior specifies how a skill is designed to generalize. 18 00:01:57,476 --> 00:02:05,656 This can be implemented in the skill through its representation or training objective, and it shapes how the skill generalizes beyond its demonstrations. 19 00:02:05,712 --> 00:02:13,612 Evan: For instance, a grasp skill might be trained in an object-relative frame, which then informs where the policy should be applicable. 20 00:02:13,612 --> 00:02:21,872 During deployment, the runtime agent reads this prior and uses it to decide which policy to invoke based on the current task requirements. 21 00:02:21,936 --> 00:02:28,156 Ashley: The paper proposes a method called Agent Priors-guided Policy Learning, or APPL. 22 00:02:28,156 --> 00:02:38,116 At its core, APPL uses structural priors that help shape the trained policies and provide runtime agent information about when those policies should be applicable. 23 00:02:38,160 --> 00:02:50,710 Evan: To implement APPL, a construction agent segments complete demonstrations into reusable skills, proposes several structural priors for each skill, and trains one documented policy per prior. 24 00:02:50,710 --> 00:02:56,660 These policies are then verified from demonstrated entry states and stored in a frozen skill library. 25 00:02:56,712 --> 00:03:07,892 Ashley: At runtime, a separate agent reads these interfaces to choose among the frozen policies, instantiate their arguments and stopping conditions, and compose them toward new task goals. 26 00:03:07,892 --> 00:03:15,352 This way, information that shaped a policy during learning remains available when that policy is later selected and composed. 27 00:03:15,408 --> 00:03:23,418 Evan: The effectiveness of APPL is demonstrated across a range of tasks from MetaWorld and long-horizon ManiSkill tasks. 28 00:03:23,418 --> 00:03:30,688 Results show improvements in out-of-distribution skill generalization and enable previously unseen skill compositions. 29 00:03:30,744 --> 00:03:37,694 Ashley: Furthermore, the paper discusses ablating interface information, which substantially reduces performance. 30 00:03:37,694 --> 00:03:45,404 These results support the use of training-time structural assumptions as a bridge between skill learning and skill composition. 31 00:03:45,516 --> 00:03:48,456 Evan: And that's the end of the Introduction section of the paper. 32 00:03:49,706 --> 00:03:53,596 Evan: Alright, let's dive into the methodology section of the paper. 33 00:03:53,596 --> 00:04:03,646 The authors introduce Agent Priors-guided Policy Learning, or APPL, which fundamentally changes how policies are learned and selected during runtime. 34 00:04:03,698 --> 00:04:10,038 Ashley: To begin with, APPL relies on two main agents: the construction agent and the runtime agent. 35 00:04:10,038 --> 00:04:20,278 The construction agent operates offline and has a series of tasks including segmenting demonstrations, proposing priors, training policies, and verifying them. 36 00:04:20,330 --> 00:04:26,280 Evan: Right, the construction agent starts by segmenting complete demonstrations into reusable skills. 37 00:04:26,280 --> 00:04:31,280 It examines each trajectory and finds distinct physical responsibilities within it. 38 00:04:31,280 --> 00:04:37,600 Then, it intentionally overlaps adjacent skill segments around their transitions or handoffs. 39 00:04:37,600 --> 00:04:46,570 This overlapping includes part of the end of its predecessor and the beginning of its successor, broadening the training support around these handoff states. 40 00:04:46,634 --> 00:04:55,774 Ashley: Which helps because the successor policy can now take over from states that the predecessor skill actually reaches, including states before nominal completion. 41 00:04:55,774 --> 00:05:04,054 It's important to note that this overlap reuses transitions from the original demonstrations and introduces no additional demonstration data. 42 00:05:04,166 --> 00:05:10,676 Evan: After segmenting demonstrations, the construction agent proposes several structural priors for each skill. 43 00:05:10,676 --> 00:05:17,886 The idea is not to commit to one predefined abstraction or policy implementation but rather to have a variety. 44 00:05:17,886 --> 00:05:26,046 Each prior encodes assumptions about relations the skill should depend on and how its behavior should respond to different changes in the scene. 45 00:05:26,090 --> 00:05:38,820 Ashley: For example, an object-relative prior expresses the assumption that a manipulation behavior can be reused at different absolute object locations when the relevant object-relative geometry is preserved. 46 00:05:38,820 --> 00:05:43,090 This helps the skill policy generalize beyond its demonstrations. 47 00:05:43,184 --> 00:05:49,264 Evan: Each proposed prior has two realizations: the training realization and the description realization. 48 00:05:49,264 --> 00:05:52,384 The training realization shapes the learned policy. 49 00:05:52,384 --> 00:06:00,854 This could involve altering the policy representation, action parameterization, architecture, or introducing auxiliary training objectives. 50 00:06:00,914 --> 00:06:14,694 Ashley: For instance, a policy class Θρi defined by a specific structural prior might be trained through the following loss function: the classic behavior cloning loss plus a regularization term derived from the prior itself. 51 00:06:14,694 --> 00:06:19,274 The prior shapes what the policy learns and consequently its success region. 52 00:06:19,322 --> 00:06:23,442 Evan: But how does the runtime agent utilize these priors? 53 00:06:23,528 --> 00:06:24,598 Ashley: Good question. 54 00:06:24,598 --> 00:06:28,558 The same prior also forms part of the policy's runtime description. 55 00:06:28,558 --> 00:06:36,298 This description communicates the hypothesized applicability region of the policy based on structural assumptions and observed training support. 56 00:06:36,298 --> 00:06:41,968 The runtime agent uses this information to select the appropriate policies during task execution. 57 00:06:41,968 --> 00:06:48,038 Thus, the prior helps improve the interface faithfulness since it is based on assumptions that shaped the policy. 58 00:06:48,098 --> 00:06:53,468 Evan: The construction agent also systematically verifies the trained policies before freezing them. 59 00:06:53,468 --> 00:07:05,458 Verification involves testing the policy on demonstrated entry states, executing each conditional diffusion policy for a set duration, and evaluating it against the demonstrated exit conditions. 60 00:07:05,552 --> 00:07:13,402 Ashley: What's crucial here is that the verification step uses limited execution evidence coming only from the demonstrated skill-entry states. 61 00:07:13,402 --> 00:07:19,982 The agent then writes a verification report, separating observed outcomes from its inferences about the policy. 62 00:07:20,102 --> 00:07:33,872 Evan: After verification, the construction agent records an interface for each trained skill policy, comprising four fields: the prior description, handoff description, support description, and verification evidence. 63 00:07:33,872 --> 00:07:40,752 These descriptions give the runtime agent the information it needs about the policy’s applicability and training support. 64 00:07:40,752 --> 00:07:45,902 Additionally, each interface declares the typed arguments required by the policy. 65 00:07:45,962 --> 00:07:50,442 Ashley: Once the library is constructed and frozen, the runtime agent takes over. 66 00:07:50,442 --> 00:07:58,192 The runtime agent receives the task goal, the current state, recent interaction history, and the available policy interfaces. 67 00:07:58,192 --> 00:08:06,522 Using this information, it chooses a policy, specifies the arguments, sets an execution duration, and defines stop conditions. 68 00:08:06,578 --> 00:08:16,768 Evan: If I understand correctly, during task execution, the executor runs the selected policy with fresh observations, checking the stop conditions after every step. 69 00:08:16,768 --> 00:08:25,598 When a stop condition is met, control returns to the runtime agent, which informs the next steps based on the updated state and goal predicates. 70 00:08:25,658 --> 00:08:26,878 Ashley: Exactly. 71 00:08:26,878 --> 00:08:35,668 This enables the runtime agent to continue the current skill, switch policies for the same skill, invoke a different skill, or terminate the task. 72 00:08:35,668 --> 00:08:45,798 Importantly, it can choose among alternative implementations for the same skill based on prior information, therefore leveraging the skill library's learned coverage in new tasks. 73 00:08:45,842 --> 00:08:52,652 Evan: This mechanism is particularly effective for handling new compositions and object shifts beyond the demonstrations. 74 00:08:52,652 --> 00:09:02,882 APPL demonstrated superior performance on MetaWorld and ManiSkill tasks compared to conventional full-task policies, providing robust skill generalization. 75 00:09:02,930 --> 00:09:08,110 Ashley: Right, and the authors also conducted experiments to test both roles of structural priors. 76 00:09:08,110 --> 00:09:15,460 One experiment evaluated the prior-designed policies on out-of-distribution states, showing substantial improvements. 77 00:09:15,460 --> 00:09:25,790 Another experiment tested the runtime agent’s ability to compose these policies in long-horizon tasks under shifted objects and task variants, again showing effectiveness. 78 00:09:25,850 --> 00:09:39,450 Evan: The results indicate that using structural assumptions both for learning and as runtime information significantly boosts performance, and hiding some structural information at runtime substantially reduces success. 79 00:09:39,506 --> 00:09:41,646 Ashley: That's the end of the Method section. 80 00:09:41,646 --> 00:09:45,146 Up next, we'll discuss the Results section of the paper. 81 00:09:46,395 --> 00:09:52,395 Evan: Alright, let's move on to the experiments and results section, which is quite comprehensive in the paper. 82 00:09:52,443 --> 00:10:01,193 Ashley: The authors conducted two primary experiments to evaluate the effectiveness of Agent Priors-guided Policy Learning, or APPL. 83 00:10:01,193 --> 00:10:09,343 The first experiment focused on skill generalization, while the second examined the composition of those skills in long-horizon tasks. 84 00:10:09,447 --> 00:10:11,957 Evan: Let’s start with the first experiment. 85 00:10:11,957 --> 00:10:14,467 What did they aim to investigate here? 86 00:10:14,523 --> 00:10:19,043 Ashley: Experiment 1 aimed to evaluate the training realization of priors. 87 00:10:19,043 --> 00:10:27,203 Specifically, it asked whether an agent can design and implement priors that make a skill’s policy generalize from a few demonstrations. 88 00:10:27,203 --> 00:10:35,223 The tasks were adapted versions of six MetaWorld tasks, including pick-place-wall, assembly, and drawer-open, among others. 89 00:10:35,283 --> 00:10:37,783 Evan: And how were these tasks set up for testing? 90 00:10:37,857 --> 00:10:41,127 Ashley: Each task had 20 successful demonstrations. 91 00:10:41,127 --> 00:10:58,507 The test states varied two spatial factors per task into three categories: IID states stayed within the demonstrated factor ranges, C recombined the factors within seen ranges, and E extrapolated both factors beyond their demonstrated intervals, hence out-of-distribution or OOD. 92 00:10:58,507 --> 00:11:03,727 What's intriguing is that some states were completely new combinations not seen during training. 93 00:11:03,771 --> 00:11:05,481 Evan: Interesting setup. 94 00:11:05,481 --> 00:11:07,831 So, what were the findings? 95 00:11:07,875 --> 00:11:17,045 Ashley: The experiment tested six different systems for each demonstration scenario, including a baseline diffusion policy and one with a fixed relational prior. 96 00:11:17,045 --> 00:11:24,795 The results showed that agent-designed priors significantly improved out-of-distribution skill generalization across all tasks. 97 00:11:24,843 --> 00:11:28,723 Evan: Could you give us some specifics? 98 00:11:28,779 --> 00:11:44,569 Ashley: For instance, with only two demonstrations, an agent’s best proposal achieved 89.6% OOD success, compared to just 28.96% for the vanilla diffusion policy and 37.92% for the fixed relational prior. 99 00:11:44,569 --> 00:11:51,979 Even with increased demonstrations, the trend remained the same, showcasing the significant impact of these agent-designed priors. 100 00:11:52,035 --> 00:11:53,175 Evan: That's impressive. 101 00:11:53,175 --> 00:11:55,455 So what about the second experiment? 102 00:11:55,515 --> 00:12:05,265 Ashley: The second experiment evaluated whether the runtime agent could effectively compose these prior-specific skill policies to complete tasks from new states and goals. 103 00:12:05,265 --> 00:12:12,305 They used five long-horizon ManiSkill tasks like drawer exchange, buffer exchange, and retrieve and store. 104 00:12:12,305 --> 00:12:15,455 Each task had twelve successful demonstrations. 105 00:12:15,507 --> 00:12:19,027 Evan: What were the methods compared in this part of the experiment? 106 00:12:19,083 --> 00:12:30,883 Ashley: They compared APPL with several baselines: a full-task diffusion policy, a single prior full-task policy, and a vision-language-action model orchestrated by the same runtime agent. 107 00:12:30,883 --> 00:12:36,223 They also tested three ablated versions of APPL that hid certain interface information. 108 00:12:36,327 --> 00:12:37,767 Evan: And what did they find? 109 00:12:37,887 --> 00:12:40,267 Ashley: APPL outperformed all baselines. 110 00:12:40,267 --> 00:12:52,627 For example, on motion-level OOD tasks, APPL achieved 50.0% success versus 10.0% for both the full-task diffusion policy and the single prior full-task policy. 111 00:12:52,627 --> 00:13:01,527 This indicates that APPL’s method of using various skill policies with runtime agent selection is quite effective for out-of-distribution generalization. 112 00:13:01,647 --> 00:13:03,627 Evan: Interesting. 113 00:13:03,627 --> 00:13:11,767 Did the results include any insights into handling the complexity of long-horizon tasks? 114 00:13:11,811 --> 00:13:13,561 Ashley: Yes, absolutely. 115 00:13:13,561 --> 00:13:18,281 APPL successfully completed various task-level OOD and composition cases. 116 00:13:18,281 --> 00:13:24,551 It managed to achieve its tasks despite variations like differing starting conditions and intermediate sub-goals. 117 00:13:24,551 --> 00:13:29,081 When the prior-related information was hidden, performance dropped significantly. 118 00:13:29,081 --> 00:13:38,711 For instance, success in task-level OOD cases dropped from 92.5% to 65.0% when prior information was hidden. 119 00:13:38,793 --> 00:13:48,063 Evan: In essence, the results clearly show that structural assumptions used both in training and as part of the runtime interface greatly improve performance. 120 00:13:48,063 --> 00:13:54,143 The experiments highlight that APPL can effectively bridge the gap between skill learning and composition. 121 00:13:54,195 --> 00:14:02,015 Ashley: Yes, the experiments also underscore the importance of having detailed structural information accessible to the runtime agent. 122 00:14:02,015 --> 00:14:09,675 By making this information available, APPL was able to more effectively generalize skills and handle new task variations. 123 00:14:09,723 --> 00:14:13,403 Evan: And that sums up the experiment and results section of the paper. 124 00:14:14,669 --> 00:14:21,209 Evan: Let's now explore how this work fits within the broader landscape of research in this field. 125 00:14:21,269 --> 00:14:29,559 Ashley: The authors categorize the related work into a few key areas, emphasizing how their approach addresses some of the limitations of existing methods. 126 00:14:29,559 --> 00:14:34,509 To start, they discuss structural priors in the context of skill generalization. 127 00:14:34,565 --> 00:14:42,585 Evan: Structural priors, such as object-relative frames, are effective in making learned behaviors adaptable to new situations. 128 00:14:42,585 --> 00:14:49,365 However, most existing works rely on priors that are fixed or designed manually for specific tasks. 129 00:14:49,421 --> 00:14:53,161 Ashley: Yes, and the authors reference studies like those by Eisner et al. 130 00:14:53,161 --> 00:14:55,681 in 2022 and Bahety et al. 131 00:14:55,681 --> 00:15:02,161 in 2024, which focus on specific operations like doors and drawers or screw motions. 132 00:15:02,161 --> 00:15:09,521 These studies show that while these priors are effective, their applicability is limited to the specific tasks they were designed for. 133 00:15:09,641 --> 00:15:16,951 Evan: Interestingly, the paper also points out that even within a task, the right structural prior can vary significantly. 134 00:15:16,951 --> 00:15:26,741 For instance, benefits from priors, like wrist-camera views or 3D point clouds, are highly task-dependent, as shown in studies by Hsu et al. 135 00:15:26,741 --> 00:15:28,881 in 2022 and Ling et al. 136 00:15:28,881 --> 00:15:30,561 in 2023. 137 00:15:30,635 --> 00:15:35,795 Ashley: Moreover, the assumptions behind these priors can limit the achievable accuracy. 138 00:15:35,795 --> 00:15:41,745 For example, an incorrect symmetry assumption might hinder performance, as discussed by Wang et al. 139 00:15:41,745 --> 00:15:43,325 in 2023. 140 00:15:43,373 --> 00:15:47,543 Evan: That's where APPL's flexibility becomes a significant advantage. 141 00:15:47,543 --> 00:15:55,133 By proposing multiple priors for each skill and letting the runtime agent choose among them, it addresses this limitation. 142 00:15:55,211 --> 00:15:59,111 Ashley: Next, the authors talk about agents and robot tools. 143 00:15:59,111 --> 00:16:09,361 They note that interfaces for these tools often include geometric feasibility checks and predicate invention for learned skills, as observed in studies like those by Lin et al. 144 00:16:09,361 --> 00:16:11,571 in 2023 and Yang et al. 145 00:16:11,571 --> 00:16:13,281 in 2025. 146 00:16:13,325 --> 00:16:19,925 Evan: These checks are essential but don't always capture the complete context required for effective skill execution. 147 00:16:19,925 --> 00:16:26,965 APPL enhances this by including handoff overlaps and verification reports, which provide richer context. 148 00:16:27,029 --> 00:16:32,559 Ashley: They also mention symbolic abstractions and skill discovery as another critical area. 149 00:16:32,559 --> 00:16:41,269 Traditional methods like options and sampler-based planning provide foundational concepts for temporal abstraction and task-and-motion planning. 150 00:16:41,333 --> 00:16:42,253 Evan: Right. 151 00:16:42,253 --> 00:16:45,883 For example, methods proposed by Konidaris et al. 152 00:16:45,883 --> 00:16:48,153 in 2018 and Garrett et al. 153 00:16:48,153 --> 00:16:54,793 in 2020 involve learning symbolic representations from skills, which can be quite powerful. 154 00:16:54,845 --> 00:16:55,845 Ashley: Exactly. 155 00:16:55,845 --> 00:17:07,685 However, the APPL approach extends these ideas by employing structural priors that inform both policy learning and runtime decision-making, a dual role that most traditional methods don't cover. 156 00:17:07,783 --> 00:17:16,723 Evan: In the same vein, the authors draw contrasts with systems that sequence learned or human-guided skills using planners or language models. 157 00:17:16,723 --> 00:17:19,153 Studies like those by Mandlekar et al. 158 00:17:19,153 --> 00:17:21,173 in 2023 and Dalal et al. 159 00:17:21,173 --> 00:17:24,353 in 2024 illustrate these approaches. 160 00:17:24,413 --> 00:17:32,013 Ashley: Yes, and importantly, transition policies and skill-chaining methods aim to manage handoffs between skills effectively. 161 00:17:32,013 --> 00:17:40,853 APPL tackles this by using overlapping training segments, which broaden the initiation set for the successor skill, making transitions smoother. 162 00:17:40,961 --> 00:17:41,891 Evan: Interesting. 163 00:17:41,891 --> 00:17:52,481 So, APPL essentially integrates ideas from various existing methods and builds upon them to create a more robust framework for skill generalization and composition. 164 00:17:52,541 --> 00:17:53,591 Ashley: Precisely. 165 00:17:53,591 --> 00:18:03,651 The authors also refer to agent-driven systems where models help design rewards, simulation tasks, and even sim-to-real rewards, like in studies by Xie et al. 166 00:18:03,651 --> 00:18:05,811 in 2024 and Wang et al. 167 00:18:05,811 --> 00:18:07,371 in 2024. 168 00:18:07,371 --> 00:18:11,941 APPL leverages similar concepts but in a more structured and targeted way. 169 00:18:12,005 --> 00:18:14,055 Evan: That's a lot of integration. 170 00:18:14,055 --> 00:18:20,905 How do they differentiate APPL from systems using pre-existing policies or coding agents? 171 00:18:20,957 --> 00:18:21,967 Ashley: Great point. 172 00:18:21,967 --> 00:18:28,727 The key difference is that APPL derives its documentation from the construction-time prior and handoff overlaps. 173 00:18:28,727 --> 00:18:36,857 This is before any actual tasks are executed, unlike systems that update their policy cards or skill memories based on execution. 174 00:18:36,967 --> 00:18:41,617 Evan: It sounds like APPL has a proactive approach rather than a reactive one. 175 00:18:41,669 --> 00:18:42,849 Ashley: Exactly. 176 00:18:42,849 --> 00:18:52,969 By structurally embedding knowledge at both the learning and decision-making stages, APPL ensures a more seamless integration of skills in varied and dynamic settings. 177 00:18:53,021 --> 00:18:56,061 Evan: And that wraps up the Related Work section of the paper. 178 00:18:57,318 --> 00:19:02,978 Evan: Alright, Ashley, let's summarize the key contributions and takeaways of this paper. 179 00:19:03,030 --> 00:19:09,580 Ashley: First and foremost, the paper introduces Agent Priors-guided Policy Learning, or APPL. 180 00:19:09,580 --> 00:19:20,350 This approach significantly enhances both skill generalization and compositional generalization by embedding structural priors into the learning and decision-making framework. 181 00:19:20,406 --> 00:19:26,966 Evan: The idea of using structural priors for both training and runtime interface is quite revolutionary. 182 00:19:26,966 --> 00:19:33,346 It addresses the common issue where information is lost between task-level goals and skill-level executions. 183 00:19:33,390 --> 00:19:34,440 Ashley: Exactly. 184 00:19:34,440 --> 00:19:42,310 By having a construction agent propose and implement multiple priors per skill, APPL creates a versatile skill library. 185 00:19:42,310 --> 00:19:50,610 The runtime agent can then make informed decisions on which policy to apply based on the specific task requirements and current states. 186 00:19:50,670 --> 00:19:51,450 Evan: Right. 187 00:19:51,450 --> 00:20:00,970 And the overlap in training segments allows for smoother skill transitions, making the system more robust in handling varied and out-of-distribution conditions. 188 00:20:01,014 --> 00:20:03,944 Ashley: The experimental results are compelling as well. 189 00:20:03,944 --> 00:20:14,154 APPL outperformed conventional approaches in both MetaWorld and long-horizon ManiSkill tasks, especially in out-of-distribution scenarios and new task compositions. 190 00:20:14,214 --> 00:20:27,374 Evan: The experiments clearly demonstrated that hiding interface information led to substantial performance drops, further validating the importance of accessible structural information for runtime agents. 191 00:20:27,468 --> 00:20:28,648 Ashley: Definitely. 192 00:20:28,648 --> 00:20:38,858 Overall, APPL shows how integrating structural assumptions as both a learning guide and a runtime interface can vastly improve robot learning and task execution. 193 00:20:38,970 --> 00:20:43,630 Evan: And that brings us to the end of today's episode on Daily Paper Cast. 194 00:20:43,630 --> 00:20:49,270 We hope you found this discussion on Agent Priors-guided Policy Learning insightful. 195 00:20:49,326 --> 00:20:51,636 Ashley: Yes, thank you for joining us! 196 00:20:51,636 --> 00:20:59,386 If you enjoyed this episode, be sure to tune in for our next one where we will delve into more cutting-edge research in AI and robotics. 197 00:20:59,430 --> 00:21:03,610 Evan: Don't forget to subscribe and leave a review if you liked the podcast. 198 00:21:03,610 --> 00:21:09,450 Until next time, stay curious and keep exploring the frontiers of AI research!