Technical Summary: Issue-Centric Autonomous Thinking Scheduling for Robots in Dynamic Environments

1. Problem Statement

Autonomous robots operating in open, changing, and partially observable environments face a critical limitation in existing scheduling paradigms. Conventional methods are typically task-centric or action-centric, relying on predefined goals, fixed state variables, and explicit action sequences. These approaches struggle when:

• Important issues are not fully predefined but generated dynamically from new observations.
• Issues are temporarily unexecutable but become solvable as new information arrives.
• Direct responses to observations may be incorrect without further reasoning (e.g., distinguishing between sensor failure and environmental interference).
• Robots need to recall interrupted tasks, update internal understanding, and prioritize unresolved problems over executing predefined lists.

The core challenge is to develop an internal thinking mechanism that can maintain, recall, and schedule cognitive issues based on their importance and the changing context, rather than merely executing a static list of actions.

2. Methodology: The ICATS Framework

The paper proposes ICATS (Issue-Centric Autonomous Thinking Scheduling), a framework that treats cognitive issues as the fundamental units of internal thinking and scheduling.

A. Cognitive Issue Representation

A cognitive issue is defined as an event, observation, task state, environmental change, internal doubt, or unresolved problem requiring later recall or reasoning. It is represented as a tuple:
$$M^t_i = \langle I^t_i, o^t_i, \hat{o}^t_i, \Delta^t_i, R^t_i, s^t_i, l^t_i, T^t_i \rangle$$
Where:

• $I^t_i$: The issue itself.
• $o^t_i$: Perceived result.
• $\hat{o}^t_i$: Expected result.
• $\Delta^t_i$: Perception–expectation difference.
• $R^t_i$: Related issues or experiences.
• $s^t_i$: Solving status (unsolved, partially solved, familiar, etc.).
• $l^t_i$: Learning status.
• $T^t_i$: Thinking importance.

B. Dynamic Status Maintenance

ICATS employs a dynamically extensible internal status ($\hat{S}^t$) rather than a fixed state vector. This status includes:

• Environmental factors ($\hat{E}^t$).
• Robot self-status ($\hat{B}^t$).
• Status of other agents/objects ($\hat{O}^t$).
• Relations among factors ($\hat{R}^t$).
• Plans, predictions, and reasoning results ($\hat{P}^t$).

New factors, objects, relations, or predictions observed or inferred during operation are added to this status. The robot generates up-to-date expectations based on this evolving status, allowing it to detect not only abnormal values but also newly emerging factors, changed relations, and potential risks.

C. Thinking Importance and Scheduling

Thinking importance ($T^t_i$) determines the priority of an issue. It is calculated based on:

1. Perception–Expectation Difference ($\Delta^t_i$): The gap between current perception and the up-to-date expectation.
2. Unresolved Degree ($u^t_i$): How far the issue is from being solved.
3. Novelty ($n^t_i$): Whether the issue is unfamiliar.
4. Learning Value ($q^t_i$): Potential for future knowledge gain.
5. Expected Future Influence ($e^t_i$): Potential for causing significant consequences (e.g., safety risks, task failure).

A thinking module (e.g., an LLM or reasoning model) can be used to judge these factors, particularly for issues that may lead to fatal results.

Scheduling Mechanism:
Issues are scheduled via a score $S^t_j = T^t_j + \lambda_u u^t_j + \lambda_q q^t_j - \lambda_c c^t_j$. Scheduling occurs in two modes:

• Free-Time Scheduling: Autonomous recall of unresolved or high-value issues when resources are available.
• Triggered Scheduling: Immediate processing when an unexpected situation (e.g., a pedestrian crossing path) deviates from expectations.

D. Processing and Learning

Once scheduled, an issue is processed to solve it, learn from it, or update internal models.

• Solving: The robot retrieves importance-guided past experiences to generate candidate solutions.
• Learning:
  • Semantic Memory: For rare or one-off issues, the robot stores a meaningful summary.
  • Neural Models: For repeated issues, data is used to train sequence models (e.g., LSTMs, Transformers).
• Feedback Loop: After processing, the thinking importance is updated. If an issue becomes familiar or solved, its importance decreases; if it remains unresolved, it retains high priority for future scheduling.

3. Key Contributions

1. Issue-Centric Scheduling: Shifts the scheduling unit from predefined actions/tasks to cognitive issues, enabling the autonomous recall and processing of unfinished, unfamiliar, changed, or high-impact issues.
2. Multi-Factor Thinking Importance: Formulates importance based on perception–expectation difference, unresolved degree, novelty, learning value, and future influence. It explicitly allows the use of LLMs to judge importance for critical issues.
3. Dynamically Extensible Status Maintenance: Introduces a mechanism where the internal status is not fixed but expands during operation to include new factors, objects, relations, and predictions, facilitating up-to-date expectation generation.

4. Experimental Results

The paper presents two controlled verification experiments:

Experiment 1: Time-Constrained Dynamic Issue Scheduling

• Setup: Compared ICATS against chronological, deadline-first, executable-first, shortest-duration, and random scheduling baselines.
• Results:
  • ICATS achieved the highest weighted reward (18.395), outperforming the best fixed baseline (Random, 15.967) by ~15.2%.
  • It achieved a dynamic-issue completion rate of 0.583, higher than all fixed baselines (max 0.533).
  • It demonstrated learning capability, successfully transferring learned rules (e.g., prioritizing narrow-window issues) to evaluation episodes, whereas fixed baselines had zero learning success.

Experiment 2: Dynamic Status Maintenance and Difference Detection

• Setup: Evaluated the ability to detect differences (e.g., temperature anomalies, pedestrian risks) using varying levels of status maintenance (Observation-only, Environment-only, Environment+Self+Objects, and ICATS-full-status).
• Results:
  • ICATS-full-status achieved a perfect F1-score of 1.000, significantly outperforming the best partial-status baseline (0.618).
  • It reduced the False Negative Rate to 0.000 (from 0.483 in partial baselines) and False Positive Rate to 0.000.
  • It enabled early detection (3.235 time units ahead) of risks based on inferred relations, a capability absent in other methods.
  • The results highlight that maintaining relations and predictions is crucial for accurate difference detection in dynamic scenarios.

5. Significance and Claims

The paper claims that ICATS provides a necessary evolution for autonomous robots in open environments by:

• Prioritizing Cognitive Needs: Moving beyond simple task execution to managing the robot's internal cognitive state, ensuring that unresolved, novel, or high-risk issues are addressed even if they are not part of the original task list.
• Adaptability: The dynamically extensible status allows robots to handle factors not predefined at the start of a mission, such as new objects or changed relations.
• Improved Safety and Efficiency: By detecting risks earlier (via relation/prediction maintenance) and prioritizing high-impact issues, the framework improves safety and resource allocation.

The authors modestly note that the verification results are from controlled experiments. While ICATS achieved perfect classification in the specific scenarios tested, the authors acknowledge that future work must evaluate robustness against noisy observations, incomplete predictions, and more complex real-world environments. The paper positions ICATS not as a replacement for action planning, but as a complementary thinking scheduler that operates before, during, or after action planning to ensure the robot is thinking about the right things at the right time.