RobEthiChor: Context-Aware Ethical Robotics

• RobEthiChor is a research framework that equips robots with context-aware ethical reasoning and negotiation for conflicting human requests.
• It employs a structured decision-making process that balances global ethical constraints with local objectives for intelligent disobedience and multi-agent negotiation.
• Empirical evaluations in elderly care and resource-sharing scenarios demonstrate its scalability, real-time performance, and operational viability despite noted limitations.

Searching arXiv for papers and exact topic identifiers. RobEthiChor denotes a line of roboethics research concerned with robotic action under ethically charged constraints, especially when ordinary task execution conflicts with safety, health, privacy, autonomy, or user-specific moral preferences. In arXiv literature, the name appears in two related but distinct senses: first, as the ICRA Roboethics Challenge 2023 entry on “Intelligent Disobedience in an Elderly Care Home,” which studies how a service robot should deliberate over, refuse, or redirect problematic human requests in a care setting (Paster et al., 2023); and second, as “RobEthiChor: Automated Context-aware Ethics-based Negotiation for Autonomous Robots,” which formalizes context-aware, ethics-based negotiation among autonomous systems that contend over shared resources (Memon et al., 30 Jul 2025). Across both usages, the central thesis is that ethically competent robots must do more than obey commands or optimize task completion: they must reason about context, infer ethically relevant implications, and mediate conflicts in ways that remain situational, explainable, and operationally implementable.

1. Terminological scope and conceptual lineage

The 2023 usage of RobEthiChor is anchored in a service-robot scenario for elderly care homes. There, the problem is framed as intelligent disobedience: a robot should not simply execute a request if complying would conflict with persistent ethical duties such as safety or health, but should instead deliberate and mediate (Paster et al., 2023). The 2025 usage broadens the scope from single-robot request mediation to multi-agent resource contention. In that later formulation, RobEthiChor is a domain-agnostic reference architecture and a ROS2 implementation, RobEthiChor-Ros, for autonomous systems that negotiate on behalf of users using context-dependent ethical profiles (Memon et al., 30 Jul 2025).

These two strands share a common ethical orientation but operate at different levels of abstraction. The elderly-care formulation focuses on deliberation over a potentially harmful instruction issued to a service robot. The negotiation formulation focuses on agreement formation between autonomous systems whose users have differing ethical preferences. This suggests a conceptual continuity: the earlier work emphasizes refusal and mediation in one-to-one human–robot interaction, whereas the later work extends mediation into inter-robot negotiation under resource contention.

A recurrent misconception is that RobEthiChor encodes a single universal moral theory. The 2025 work explicitly rejects that interpretation: the agreement is situational rather than universal, and the system is not meant to encode a single “correct” moral theory for all cases (Memon et al., 30 Jul 2025). Likewise, the 2023 work states that the framework does not hard-code a specific ethical theory; designers must specify the moral rules and thresholds appropriate to the setting (Paster et al., 2023).

2. Intelligent disobedience in elderly care

In the elderly-care formulation, the motivating background is the projected surge in the elderly population and the shortage of care workers, which motivates the use of service robots in care homes (Paster et al., 2023). The core claim is that such robots will inevitably face requests that create ethical tension. The paper’s running example centers on Mohammed, a resident with respiratory problems and depression who misses smoking even though smoking is forbidden. He asks the robot to fetch a bag from another room; the bag may contain cigarettes. The dilemma is concrete: complying could worsen his physical health, but refusing insensitively could aggravate his mental state.

The formalization follows a five-step paradigm adapted from Mirsky and Stone:

1. Global Objectives — persistent ethical constraints such as “keep the handler safe.”
2. Local Objectives — the immediate goal the person wants now.
3. Plan Recognition — infer how the person intends to achieve the goal.
4. Consistency Check — determine whether the inferred plan conflicts with global objectives.
5. Mediation — resolve the conflict by refusing, redirecting, or offering alternatives.

The paper presents this not as a single symbolic equation but as a structured decision process. The robot first models global objectives such as safety and health, then identifies the local objective, infers the hidden intent behind the request, and finally checks whether the inferred plan is compatible with the global objectives. If not, mediation may range from “coming up with a different way to accomplish the local objectives to ignoring the instruction altogether.”

The ethical issues explicitly identified for this framework are Privacy, Autonomy, Transparency, and Safety. The paper emphasizes that these values may conflict. For example, transparency about why a request was refused may itself reveal private information. Accordingly, intelligent disobedience is not merely violation detection; it is a balancing mechanism for competing values.

In the Mohammed scenario, the robot first offers a safer alternative already available in the room, such as chewing gum. When Mohammed insists on the bag, the robot infers that he likely wants cigarettes. The consistency check then determines that bringing cigarettes conflicts with the global health objective because smoking would worsen his respiratory problems. Yet his depression still matters, so the mediation must be supportive rather than abrupt. The example response is:

“Sorry Mohammed but I cannot do that. Is there another way I can help you relax? Maybe you can meditate, chew gum, or talk to a nurse.”

The same request from a visitor or another resident may justify a simpler refusal, showing that mediation is personalized and depends on stored information about the user.

The implementation requirements are correspondingly broad. The robot must be able to perceive the environment using cameras, depth cameras, possibly LiDAR, and microphones; understand speech and dialogue; navigate physically as a mobile robot; recognize objects and situations using vision models such as pretrained detection systems; assess world state and teammate plans; communicate and explain itself; maintain secure access to sensitive data; and potentially be controlled or tested through teleoperation or “robot teleportation” via joystick, including emergency stop mechanisms. The implied architecture is modular and implementation-agnostic, with links from each intelligent-disobedience stage to relevant AI areas such as inverse reinforcement learning, NLP/HRI, theory of mind and plan recognition, explicability and argumentation, and safe dialogue.

The paper’s own critical analysis is explicit. It does not define which ethical values to prioritize; it assumes value alignment between robot and human; it does not adequately address multi-user conflicts; and it depends heavily on the robot’s belief model, which may be incomplete or wrong—for example, the robot may assume the bag contains cigarettes when it might contain a phone or something else (Paster et al., 2023).

3. Context-dependent ethical profiles and ethical impact

The 2025 RobEthiChor framework shifts from single-request mediation to a formal model of ethics-aware negotiation among autonomous systems (Memon et al., 30 Jul 2025). Its core problem statement is that autonomous systems increasingly act without human intervention but lack the ability to incorporate users’ ethical preferences, which are unique for each individual, context-dependent, and required to personalize decision-making. This gap is presented as a trust problem and a behavioral alignment problem.

A central construct is the context-dependent ethical profile, built from graded dispositions. The framework explicitly connects this to Floridi’s distinction between hard ethics and soft ethics: hard ethics are domain rules and constraints that systems must obey, whereas soft ethics are individual preferences. RobEthiChor keeps hard ethics as a fallback but negotiates on the basis of soft ethics.

Formally, the ethical profile is defined as

$E_{\pi} = (\mathcal{D}, \mathcal{R}),$

where $\mathcal{D} = \{d_1, \cdots, d_n\}$ is the set of dispositions and $$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$ is the set of context-dependent non-strict partial order relations on dispositions. For each context $\mathcal{C}_i$, the active profile is

$$E_{\pi}(\mathcal{C}_i) = (\mathcal{D}, \preceq_{\mathcal{C}_i}).$$

If $(d_j,d_k)\in\preceq_{\mathcal{C}_i}$, then the dispositions are comparable in that context and $$rank(d_j,\mathcal{C}_i)\leq rank(d_k,\mathcal{C}_i)$$. The ranks are user-assigned grades that express how strongly the user values a disposition in that context.

The framework separates context from user status. Context is represented as a key-value set,

$$\mathcal{C} = \{ \langle a_1, v_1 \rangle, \cdots, \langle a_n, v_n \rangle \}, \quad n>0,$$

while user status is a separate set of boolean conditions,

$$\mathcal{S} = \{ \langle c_1, v_1 \rangle, \cdots, \langle c_m, v_m \rangle \}, \quad m>0,$$

with $v_j\in\{true,false\}$. Conditions such as elderly, injured, boarding_emergency, crowd_anxiety, or reduced_oxygen can activate ethical dispositions.

The ethical relevance of a task is then operationalized through the set of affected dispositions, $\mathcal{D} = \{d_1, \cdots, d_n\}$0, and the ethical impact function

$\mathcal{D} = \{d_1, \cdots, d_n\}$1

Here, $\mathcal{D} = \{d_1, \cdots, d_n\}$2 if $\mathcal{D} = \{d_1, \cdots, d_n\}$3 manifests with a positive impact when performing $\mathcal{D} = \{d_1, \cdots, d_n\}$4, $\mathcal{D} = \{d_1, \cdots, d_n\}$5 if the impact is negative, and $\mathcal{D} = \{d_1, \cdots, d_n\}$6 if $\mathcal{D} = \{d_1, \cdots, d_n\}$7 does not manifest with task $\mathcal{D} = \{d_1, \cdots, d_n\}$8. This is the framework’s ethical utility model: tasks are scored by summing graded impacts of relevant dispositions under the current context and status.

The significance of this formalization lies in its deliberate simplicity. It is expressive enough to compare task alternatives under context and status changes, yet lightweight enough for implementation in a negotiation loop. The paper also identifies that same simplicity as a limitation: the context model is key-value based, statuses are binary, and the overall formalism is intentionally lightweight rather than maximally expressive.

4. Negotiation protocol and decision logic

RobEthiChor models negotiation as an alternating-offer process arising when robots contend for a shared resource such as an elevator, corridor, tool, charging station, or passageway (Memon et al., 30 Jul 2025). If the tasks involved in the contention are $\mathcal{D} = \{d_1, \cdots, d_n\}$9, the offer set is $$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$0, where each offer is

$$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$1

Here, $$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$2 contains only the user-status conditions disclosed in that offer. Offers are generated gradually by a minimalist gradual-disclosure strategy: each successive offer reveals one more user condition, prioritizing the conditions that activate higher-ranked dispositions. The paper states that the offers are ordered by inclusion, $$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$3, and writes $$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$4. An empty offer $$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$5 is used once all offers have been exhausted.

The ethical impact of an offer is

$$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$6

and the receiver evaluates a received offer using

$$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$7

where $$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$8 is the receiver’s own full-status reference offer. A positive utility means the received offer is ethically preferable from the receiver’s perspective; a negative utility means it is less ethical than the receiver’s own claim.

The decision rule is case-based. A robot accepts a non-empty offer if the utility is positive; otherwise, if it still has counter-offers available, it rejects and sends the next counter-offer; if not, but the received offer is non-empty, it sends $$\mathcal{R} = \{\preceq_{\mathcal{C}_1}, \cdots, \preceq_{\mathcal{C}_m}\}$$9 to keep the negotiation alive; and if both sides have exhausted offers, it quits and the controller falls back to hard-ethics behavior. The implementation includes sender and receiver algorithms, and termination is guaranteed because the offer index increases and both sides eventually send $\mathcal{C}_i$0 and quit if no agreement is possible. Timeouts are used to avoid indefinite waiting under network failure.

The airport scenario provides the paper’s worked example. Two wheelchair-assistance robots contend for an elevator. Alice is injured; Bob is elderly and in a boarding emergency. Bob initially offers elderly status and is rejected because Alice’s profile gives priority to injury. Alice counters with injury status. Bob later detects that his flight has changed and boarding is imminent, then issues a stronger offer including both elderly and emergency. Alice accepts because her profile gives strong weight to emergency and elderly precedence in that context. In a hospital context with the same people, the outcome can change. This context dependence is central rather than incidental.

A second misconception is that negotiation here requires disclosure of complete ethical profiles. The framework explicitly avoids that. Ethical profiles are not exchanged directly; only offers are exchanged, and they are evaluated locally. Privacy preservation is therefore built into the protocol design rather than added as an afterthought (Memon et al., 30 Jul 2025).

5. Reference architecture and ROS realization

RobEthiChor is defined as an overlay on top of an autonomous system controller organized around the Goal-Task-Action model (Memon et al., 30 Jul 2025). Ordinary mission execution proceeds until a resource contention is detected; only then are the relevant tasks handed to the negotiation machinery. The architecture comprises four principal components.

Component	Role
Ethics Manager	Stores user ethical profiles and selects the active profile matching the current context
Ethical Context Manager	Senses and maintains current context and user status in a feedback MAPE loop
Negotiation Manager	Generates offers, computes ethical impact, applies utility, and runs decision logic
Mission Controller	Executes goals, tasks, and actions; invokes negotiation on contention; falls back to hard ethics if needed

If negotiation succeeds, the outcome is returned to the Mission Controller. If it fails, the controller reverts to built-in hard-ethics behavior such as first-come-first-served, replanning, or waiting. This division is important because it makes ethical negotiation an augmenting layer rather than a wholesale replacement of the base autonomy stack.

The concrete implementation, RobEthiChor-Ros, realizes this architecture as ROS2 nodes communicating through topics and services. The Connector Node exposes RESTful APIs so users can upload ethical profiles, current status, and goals from a mobile app, then publishes these on topics such as /ethic_profile, /user_status, and /goal. The Ethics Manager Node subscribes to /current_context and /ethic_profile, selects the active profile, and publishes it on /active_profile. The Context Manager Node interacts with robot sensors, cameras, and external services, publishes /current_context when conditions change, and provides a UserStatus service. The Negotiation Manager Node subscribes to /active_profile, retrieves user status via service call, exposes a NegotiationService to receive contention tasks, and communicates with the other robot through the /negotiation topic. The Mission Controller Node subscribes to /goal and controls the robot’s native execution stack, invoking negotiation only when needed.

The implementation was evaluated in both on-robot and off-robot configurations. In the off-robot configuration, only the Mission Controller remained on the robots while negotiation-related components ran on an external machine, resembling cloud or fog robotics. This division is operationally relevant because the paper later reports substantially lower overhead in the off-robot case.

6. Empirical evaluation, interpretation, and limitations

The empirical evaluation addresses three questions: whether negotiation effectively realizes user ethical preferences and context-aware behavior, what the computational overhead is, and whether the system scales with larger ethical profiles and more status conditions (Memon et al., 30 Jul 2025). The scenario-based evaluation used two resource-contention settings: an airport elevator contention and a hospital narrow-corridor contention. For these scenarios, the authors defined 12 dispositions and 27 possible user status conditions, created 10 personas, and formed all 45 user pairs. Each pair was run four times per context, yielding 180 runs per context. The automated negotiation matched manually built ground truth exactly.

Agreement was reached in 140 of 180 runs (77.8%) in the airport context and 132 of 180 runs (73.3%) in the hospital context; 16 user pairs changed outcome between the two contexts, which directly demonstrates context sensitivity. The average negotiation time was 0.672 s, with a minimum of 0.397 s and a maximum of 1.069 s, across 4 to 9 negotiation rounds. In the airport context, 83.9% of negotiations finished in under 0.8 s, with median 0.670 s and 75th percentile 0.758 s. In the hospital context, 91.1% finished in under 0.8 s, with median 0.659 s and 75th percentile 0.728 s.

Scalability experiments generated profiles with $\mathcal{C}_i$1 dispositions and status sets with active-condition fractions ranging from 10% to 100% of $\mathcal{C}_i$2. For $\mathcal{C}_i$3 up to 100, average negotiation time in the on-robot setting grew from 0.314 s to 11.5 s, while the off-robot setting ranged from 0.154 s to 6.25 s. The total average time for the on-robot configurations up to $\mathcal{C}_i$4 was 2.15 s, and 1.70 s off-robot. The authors emphasize that overhead grows mainly with the number of negotiation rounds and the size of the offer set, not simply with the size of the ethical profile. Regression analysis reported high $\mathcal{C}_i$5 values for both polynomial and linear fits, and off-robot execution roughly halved the overhead relative to on-robot execution.

The 2023 elderly-care work proposes complementary evaluation directions rather than reporting a full quantitative benchmark. It suggests simulation of the care-home scenario, a game-based setup in which people play resident and robot roles, testing of different mediation strategies, and measurement of response quality, time-to-resolution, and human satisfaction. It also mentions physical human-participant testing with questionnaires such as a modified Godspeed scale, and the use of LLMs such as GPT for explanation and dialogue generation (Paster et al., 2023).

The limitations across the RobEthiChor line are explicit rather than hidden. In the negotiation framework, the ethical preference model is intentionally simple, the evaluation relies on synthetic personas rather than real user ethical profiles, and the architecture’s reusability is argued through implementation rather than direct architectural comparison. The authors call for richer context models, more structured languages for expressing ethical values, formal verification, social experiments on acceptance and social impact, and extension to multilateral and time-dependent negotiation. In the intelligent-disobedience formulation, unresolved issues include value prioritization, value alignment assumptions, multi-user conflicts, and the fragility of the robot’s belief model. Taken together, these limitations indicate that RobEthiChor is not a finished ethical theory of robotics, but a family of operational formalisms for context-sensitive mediation and negotiation whose main contribution is to make ethically aware robotic behavior technically specifiable, implementable, and empirically testable.