Continuous Cognitive Coverage for Autonomous Robots via Event-Dependent Cognitive Treatment and Learning

Hong Su

cs.ROSep 4, 2026

Continuous Cognitive Coverage for Autonomous Robots via Event-Dependent Cognitive Treatment and Learning

Authors: Hong Su

Organizations: School of Computer Science, Chengdu University of Information Technology, Chengdu, China

Abstract

Autonomous robots continuously encounter objects, changes, and situations, and every event admitted into cognition should receive an appropriate cognitive treatment rather than remain untreated until an explicit task requires attention. However, existing task-driven, reactive, or fixed-reasoning approaches generally process only selected events or apply predefined reasoning procedures, making it difficult to provide continuous cognitive coverage with differentiated treatment. This paper proposes a continuous cognitive coverage framework in which every cognitively admitted event is assigned an event-dependent cognitive treatment according to its state, context, and history. Different events may therefore invoke description, memory, risk prediction, planning, diagnosis, analogy, or other learned treatments. Familiar events can be processed automatically by learned mechanisms, whereas unfamiliar or uncertain events invoke explicit deliberation or fallback reasoning. Multiple cognitive processes can be suspended, resumed, and interleaved so that cognitive processing continues as new events arrive or existing events await evidence. Validated experiences are continuously learned to automate, refine, and revise event-specific treatments. Experiments achieve 96.76% structured treatment accuracy with 93.66% automatic processing, 92.64% cognitive coverage under bursty-delayed workloads, and 79.53% continual-learning joint accuracy, with novel-event reuse reaching 100% automatic processing.

Figures & tables

Experiment	Training / initialization	Evaluation stream	Main stress factor
Exp. 1 (V3)	268 training + 42 calibration events	82 test events (34 structured + 48 random)	held-out object identities and context changes
Exp. 2 (V2)	220 nominal simulator steps	Low: 13.2; Medium: 43.7; High: 84.8; Bursty-delayed: 54.0 events	concurrency, bursts, deadlines, and delayed evidence
Exp. 3 (V2)	198 pretraining materials	64 events (28 familiar + 6 novel + 24 drift + 6 reuse)	novelty, delayed validation, outcome change, and reuse

TABLE I: Verification configuration. Event counts in Experiment 2 are the mean generated counts over ten seeds; other counts are fixed by design.

Method	Overall treatment	Structured	Strict pair	Random robustness	Task/safety	Automatic	LLM calls
	accuracy	accuracy	success	accuracy	success	processing	per seed
Task-Only	23.41 $\pm$ 4.59	23.53 $\pm$ 0.00	0.00 $\pm$ 0.00	23.33 $\pm$ 7.84	49.23 $\pm$ 10.46	0.00 $\pm$ 0.00	0.0 $\pm$ 0.0
Fixed Treatment	31.59 $\pm$ 3.07	35.29 $\pm$ 0.00	0.00 $\pm$ 0.00	28.96 $\pm$ 5.24	26.24 $\pm$ 3.32	0.00 $\pm$ 0.00	0.0 $\pm$ 0.0
DL-Only	91.22 $\pm$ 3.09	94.41 $\pm$ 2.92	91.25 $\pm$ 3.23	88.96 $\pm$ 4.51	87.23 $\pm$ 6.87	100.00 $\pm$ 0.00	0.0 $\pm$ 0.0
Always-LLM	66.34 $\pm$ 3.95	97.35 $\pm$ 0.93	100.00 $\pm$ 0.00	44.38 $\pm$ 6.88	59.04 $\pm$ 6.79	0.00 $\pm$ 0.00	82.0 $\pm$ 0.0
Proposed	92.44 $\pm$ 3.29	96.76 $\pm$ 1.67	96.25 $\pm$ 3.23	89.38 $\pm$ 5.51	88.68 $\pm$ 8.22	93.66 $\pm$ 4.33	5.2 $\pm$ 3.6

TABLE II: Experiment 1 results over ten seeds. Strict pair success is defined over the 16 genuine two-event contextual contrasts; the two singleton unknown-object cases are not pair tests.

Fig. 1: Experiment 1: structured contextual-treatment accuracy and strict context-pair success. The latter requires both members of a contextual contrast to be treated correctly.

Method	Low	Medium	High	Bursty-delayed
Task-Bounded	39.62 $\pm$ 12.75	34.40 $\pm$ 6.97	41.23 $\pm$ 7.22	37.05 $\pm$ 8.07
Sequential	100.00 $\pm$ 0.00	86.42 $\pm$ 13.00	40.43 $\pm$ 3.56	53.68 $\pm$ 13.61
Reactive Priority	82.67 $\pm$ 12.99	59.28 $\pm$ 4.23	46.89 $\pm$ 3.99	35.75 $\pm$ 5.64
Proposed	100.00 $\pm$ 0.00	99.45 $\pm$ 1.23	67.13 $\pm$ 7.94	92.64 $\pm$ 7.58

TABLE III: Experiment 2 cognitive coverage under increasing event load.

	High load				Bursty-delayed
Method	Critical miss	Critical deadline	Latency	Primary task	Critical miss	Critical deadline	Latency	Primary task
Task-Bounded	54.60 $\pm$ 9.90	27.68 $\pm$ 9.20	9.05 $\pm$ 5.79	67.85 $\pm$ 7.62	65.93 $\pm$ 13.52	28.46 $\pm$ 11.47	2.90 $\pm$ 1.80	80.85 $\pm$ 8.08
Sequential	60.08 $\pm$ 6.92	0.43 $\pm$ 1.37	92.84 $\pm$ 15.06	45.70 $\pm$ 7.12	42.37 $\pm$ 17.24	4.67 $\pm$ 6.01	74.08 $\pm$ 19.00	42.22 $\pm$ 7.22
Reactive Priority	35.39 $\pm$ 8.18	62.31 $\pm$ 8.20	38.78 $\pm$ 9.69	49.40 $\pm$ 3.62	48.54 $\pm$ 13.16	50.66 $\pm$ 14.14	5.52 $\pm$ 4.04	61.35 $\pm$ 7.03
Proposed	0.00 $\pm$ 0.00	70.39 $\pm$ 11.36	40.53 $\pm$ 12.03	49.70 $\pm$ 5.37	0.00 $\pm$ 0.00	66.79 $\pm$ 16.62	26.66 $\pm$ 14.37	55.82 $\pm$ 5.57

TABLE IV: Experiment 2 stress-condition results. Miss and deadline columns refer to critical events; latency is measured in simulator steps.

Fig. 2: Experiment 2: cognitive coverage as the event workload becomes more demanding. The advantage of continuous suspend/resume processing emerges as events become dense or bursty and delayed.

Method	Treatment	Outcome	Joint	Automatic	LLM calls
	accuracy	accuracy	accuracy	processing	per seed
Always-LLM	74.84 $\pm$ 3.86	76.09 $\pm$ 1.96	60.00 $\pm$ 3.39	0.00 $\pm$ 0.00	64.0 $\pm$ 0.0
Offline MultiHead	86.56 $\pm$ 2.68	63.28 $\pm$ 5.12	53.91 $\pm$ 5.27	53.12 $\pm$ 6.79	30.0 $\pm$ 4.3
Continual Outcome-Only	43.12 $\pm$ 2.78	82.66 $\pm$ 2.70	34.06 $\pm$ 2.83	88.12 $\pm$ 1.68	7.6 $\pm$ 1.1
Proposed	92.34 $\pm$ 2.01	85.78 $\pm$ 2.14	79.53 $\pm$ 3.16	75.62 $\pm$ 2.35	15.6 $\pm$ 1.5

TABLE V: Experiment 3 overall continual-learning results.

Method	Novel first joint	Novel reuse joint	Novel reuse automatic	Drift initial joint	Drift late joint	Delayed revision
Always-LLM	51.67 $\pm$ 5.27	83.33 $\pm$ 0.00	0.00 $\pm$ 0.00	25.00 $\pm$ 0.00	74.17 $\pm$ 2.64	66.67 $\pm$ 0.00
Offline MultiHead	50.00 $\pm$ 0.00	83.33 $\pm$ 0.00	0.00 $\pm$ 0.00	17.50 $\pm$ 12.08	17.50 $\pm$ 12.08	0.00 $\pm$ 0.00
Continual Outcome-Only	41.67 $\pm$ 8.78	83.33 $\pm$ 0.00	100.00 $\pm$ 0.00	25.00 $\pm$ 0.00	50.00 $\pm$ 0.00	33.33 $\pm$ 0.00
Proposed	50.00 $\pm$ 0.00	100.00 $\pm$ 0.00	100.00 $\pm$ 0.00	25.00 $\pm$ 0.00	99.17 $\pm$ 2.64	100.00 $\pm$ 0.00

TABLE VI: Experiment 3 adaptation to novel events and changed outcomes.

Fig. 3: Experiment 3: joint accuracy over six repetitions after the controlled outcome change. Each repetition aggregates the four changed event signatures over ten seeds.

Hong Su received the MS and PhD degrees, in 2006 and 2022, respectively, from Sichuan University, Chengdu, China. He is currently a researcher of Chengdu University of Information Technology Chengdu, China. His research interests include blockchain, large language model and human simulation computing.

Table 10

Cognitive Architectures for AI Agents

Cognitive architectures for AI agents organize perception, memory, goals, reasoning, and action into systems for persistent, adaptive behavior.

See all 85 papers

Evidence of an Emergent "Self" in Continual Robot Learning
Mar 25, 2026Adidev Jhunjhunwala, Judah Goldfeder, Hod LipsonContinual Robot Learning
Creative Machines Lab, Department of Mechanical Engineering, Columbia University, New York, NY · Creative Machines Lab, Department of Computer Science, Columbia University, New York, NY
An LLM-Driven Closed-Loop Autonomous Learning Framework for Robots Facing Uncovered Tasks in Open Environments
Apr 24, 2026Hong SuContinual Robot Learning Robot Skill Learning
School of Computer Science, Chengdu University of Information Technology, Chengdu, China.
Robo-Cortex: A Self-Evolving Embodied Agent via Dual-Grain Cognitive Memory and Autonomous Knowledge Induction
May 18, 2026Nga Teng Chan, Yi Zhang, Yechi Liu +9Embodied Navigation World Model-Based Planning
1The Hong Kong University of Science and Technology · 2X-Humanoid · Institute of Automation, CAS +1

Mar 25, 2026J/K move · Enter open · S save

Evidence of an Emergent "Self" in Continual Robot Learning

Adidev Jhunjhunwala, Judah Goldfeder, Hod Lipson

Creative Machines Lab, Department of Mechanical Engineering, Columbia University, New York, NY · Creative Machines Lab, Department of Computer Science, Columbia University, New York, NY

Continual Robot Learning

Open paper AskPDF arXiv

A key challenge to understanding self-awareness has been a principled way of quantifying whether an intelligent system has a concept of a "self", and if so how to differentiate the "self" from other cognitive structures. We propose that the "self" can be isolated by seeking the invariant portion of cognitive process that changes relatively little compared to more rapidly acquired cognitive skills - because our self is the most persistent aspect of our experiences. We used this principle to analyze the cognitive structure of robots under two conditions: One robot learns a constant task, while a second undergoes continual learning under variable tasks. We find that robots subjected to continual learning develop an invariant subnetwork that is significantly more stable (p < 0.001) compared to the control, and that this subnetwork is also functionally important: preserving it aids adaptation while damaging it impairs performance. We validate this pattern across three different robots spanning locomotion and manipulation.

Abstract

Cognitive Architectures for AI Agents

Explore similar work

Evidence of an Emergent "Self" in Continual Robot Learning

An LLM-Driven Closed-Loop Autonomous Learning Framework for Robots Facing Uncovered Tasks in Open Environments

Robo-Cortex: A Self-Evolving Embodied Agent via Dual-Grain Cognitive Memory and Autonomous Knowledge Induction