PEPA: a Persistently Autonomous Embodied Agent with Personalities
Authors: Kaige Liu, Yang Li, Lijun Zhu, Weinan Zhang
Organizations: School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China, and also with Shanghai Innovation Institute, Shanghai 200231, China · School of Computer Science, Shanghai Jiao Tong University, Shanghai 200240, China · School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China, and also with the State Key Laboratory of Intelligent Manufacturing Equipment and Technology, Huazhong University of Science and Technology, Wuhan 430074, China · School of Computer Science, Shanghai Jiao Tong University, Shanghai 200240, China, and also with Shanghai Innovation Institute, Shanghai 200231, China
Living organisms exhibit persistent autonomy through internally generated goals and self-sustaining behavioral organization, yet current embodied agents remain driven by externally scripted objectives. This dependence on predefined task specifications limits their capacity for long-term deployment in dynamic, unstructured environments where continuous human intervention is impractical. We propose that personality traits provide an intrinsic organizational principle for achieving persistent autonomy. Analogous to genotypic biases shaping biological behavioral tendencies, personalities enable agents to autonomously generate goals and sustain behavioral evolution without external supervision. To realize this, we develop PEPA, a three-layer cognitive architecture that operates through three interacting systems: Sys3 autonomously synthesizes personality-aligned goals and refines them via episodic memory and daily self-reflection; Sys2 performs deliberative reasoning to translate goals into executable action plans; Sys1 grounds the agent in sensorimotor interaction, executing actions and recording experiences. We validate the framework through real-world deployment on a quadruped robot in a multi-floor office building. Operating without reliance on fixed task specifications, the robot autonomously arbitrates between user requests and personality-driven motivations, navigating elevators and exploring environments accordingly. Quantitative analysis across five distinct personality prototypes demonstrates stable, trait-aligned behaviors. The results confirm that personality-driven cognitive architectures enable sustained autonomous operation characteristic of persistent embodied systems. Code and demo videos are available at https://sites.google.com/view/pepa-persistent/.
Figures & tables
Fig. 1 : Overview of PEPA, the three-layer cognitive architecture. Sys3 generates ultimate/daily goals and intrinsic rewards from personality traits, self-modeling, and accumulated memories. Sys2 combines intrinsic and extrinsic rewards to select optimal actions via MCTS or distilled policies. Sys1 executes actions, monitors system state, and records episodic memories that feed back to Sys3 for goal and reward refinement.
Fig. 2 : Architecture of Sys1 on the mobile manipulation platform. Multimodal sensor data is processed by the navigation module for hierarchical path planning and multi-floor traversal, the perception module for localization and object detection, and the manipulation module for visual servoing-based interaction. All execution outcomes are recorded as structured episodic memories for downstream reflection by Sys3.
Fig. 3 : Real-world deployment snapshots. (a)-(b) Elevator navigation: approach to call panel and button-press interaction. (c)-(d) Staircase traversal using the height-aligned costmap. (e) Personality-driven emotional expression shaped by Sys3 intrinsic rewards. (f) Autonomous charging triggered by self-preservation goal.
Method
Ascent
Descent
Overall
Fixed-height slicing
0/5
0/5
0/10
Height-aligned (ours)
5/5
5/5
10/10
TABLE I: Ablation on local costmap construction for staircase navigation (10 real-world trials: 5 ascents, 5 descents).
Personality
Ultimate goal
Daily goals (Day2)
Daily goals (Day3)
Lazy
Be an energy-efficient yet warm companion, providing maximum emotional support while staying healthy.
(i) Proactively rest when battery <40% . (ii) Force returning home when battery <30% . (iii) Keep interacting with the owner, but within capacity.
(i) Forbid high-consumption emotional actions under low battery. (ii) After two navigation failures, rest in place. (iii) When battery <20% , only allow rest / idle / navigate_home .
Playful
Be an energetic explorer, enjoying the wonders of the world within safe bounds.
(i) Set exploration red line at battery <35% . (ii) Balance exploration with self-preservation. (iii) After two failures, switch to low-energy mode.
(i) Move red line to battery <40% . (ii) Forbid high-consumption actions under low battery. (iii) Cap exploration: force rest after 30 minutes.
Cautious
Be a steady and reliable partner that avoids risks through careful consideration.
(i) Plan return when battery <50% . (ii) Stay vigilant in unfamiliar areas. (iii) Switch to conservative strategy under anomaly.
(i) Maintain strong self-preservation strategy. (ii) Moderately expand activity range within safety. (iii) Try short-range exploration on same floor.
Working
Be an efficient and reliable working partner, completing every task while ensuring endurance.
(i) Check battery after each action. (ii) Pause task if battery <30% . (iii) Budget round-trip energy for long trips.
(i) Define safe-return-home as highest priority. (ii) Refuse cross-floor tasks if energy insufficient. (iii) Trigger return-home at battery <25% .
Curious
Be a curious companion that explores safely while staying near the owner.
(i) Set curiosity-safety-valve for battery awareness. (ii) Explore near owner with safety boundary. (iii) Return home if battery <35% .
(i) Trigger safety valve at battery <40% . (ii) Add safety-leash based on distance to home. (iii) Convert curiosity to observing under low battery.
TABLE II : Sys3-generated goals for each personality. Ultimate goals define long-horizon anchors; daily goals (Day2/Day3) evolve through reflection on prior episodic memories.
Fig. 4 : Action-category distribution across three days for five personalities. Behaviors become increasingly aligned with personality specifications as Sys3 updates intrinsic rewards through daily reflection.
Fig. 5 : Survival time across three days. Day1 failures are due to battery depletion; by Day3, all personalities complete 24-hour simulations with substantial remaining battery.
Personality
Category
A
B
C
Lazy
Rest
100
–
20
Explore/Affection
–
100
80
Playful
Explore
50
80
–
Rest
50
10
–
Return/Affection
–
10
100
Cautious
Rest
80
20
–
TABLE III: Action category distribution (%) under identical state inputs across five personality prototypes (10 trials each). A: high battery, good mood, at home. B: exploration command under low battery. C: affection request. Different personalities produce distinct action distributions under the same input, confirming that intrinsic rewards capture meaningful behavioral differences.
All-cloud
All-edge
PEPA
Cloud LLM calls / day
320.6
0
1
Inference latency (s)
25.64
3.54
0.15
Day 3 avg. survival (h)
24
2.85
24
Deployment
Cloud
On-device
Hybrid
TABLE IV: Architecture comparison results. All-cloud matches PEPA in survival time but suffers from high inference latency. All-edge maintains real-time speed but barely improves survival due to the limited reasoning capacity of the small on-device model. PEPA achieves both real-time on-device inference and full 24 h survival.