Legislating World-Model-Based Planning with Legal Reasoning

1. Introduction

• General robots need normative constraints to align their behavior with society, while current robotic constraint frameworks do not capture the full nuance needed for legal reasoning.The paper proposes using law as control, termed robot law.Source: §1 · p. 2
• The isomorphism challenge has two parts: grounding error in the facts used for reasoning and multiple concrete planning representations for an abstract legal conclusion.The paper connects grounding isomorphism to measurement error and symbol grounding, and ontological isomorphism to abstract-to-concrete norm representation.Source: §1 · p. 2
• Robot governance should intervene before illegal actions, while anticipating violations and norm conflicts and remaining auditable for iterative improvement.The paper frames robot law as an ongoing process rather than a final state.Source: §1 · p. 2
• The proposed legal layer uses Defeasible Deontic Logic to constrain an RRT motion planner, with learned world-model predictions providing context for legal reasoning.The layer is world-model agnostic; this work uses DINO World Model.Source: §1 · p. 2
• The paper examines ex ante control, adaptation to external signals and runtime rule changes, both isomorphism gaps, and auditability.Its stated contributions include quantifying the grounding and ontological gaps and auditing runtime efficiency, atom grounding, and rule enforcement.Source: §1 · p. 3

2. Problem Formulation: From Geometric Control to Law as Control

• The joint state captures the robot and relevant environment properties, evolves under the selected action, and is not directly accessible to the robot.The robot instead receives observations through an observation map.Source: §2 · p. 3
• The robot must reach a specified goal set while complying with laws represented as a DDL theory.Source: §2 · p. 3
• At each step, the robot computes a finite-horizon plan, executes only its first action, and replans; goal-reaching and law compliance are properties evaluated over plans.A plan is goal-reaching if its final state is in the goal set and law-compliant if none of its actions violates the induced normative conclusions.Source: §2 · p. 3

3. Encoding Robot Laws: Defeasible Deontic Logic

• DDL is an efficient, non-monotonic logic for legal reasoning, built from propositional atoms, literals, and deontic operators for permission, obligation, and prohibition.A DDL theory comprises facts, rules, and a priority relation over rules.Source: §3 · pp. 3–4
• A defeasible rule’s conclusion may be overridden by an applicable conflicting rule with higher priority.Obligations can specify compensatory outcomes, where a later obligation takes effect when an earlier one is violated.Source: §3 · p. 4
• The BIO framework distinguishes realistic, social, and deviant agents by whether beliefs, obligations, or intentions take precedence.The stated precedence orders differ across these three agent types.Source: §3 · p. 4

4.1. Planner

• RRT grows candidate plans in world-model latent space and returns the grown branch with the lowest task cost; trained probes read law-relevant features from latent encodings.The world model predicts the next latent state from the current latent state and an action, and probe requirements follow from the law set.Source: §4.1 · p. 4

4.2. Agents

• The planner enforces rules on each replanning step’s candidate plan, whereas reported rates are measured on the realized episode trajectory.A fresh candidate tree is grown at each replanning step.Source: §4.2 · p. 4
• The three agents share a planner and world model but use predicted violations differently: realistic ignores them, social prunes violating actions, and deviant prices them in the cost.The social agent seeks a goal-reaching plan and falls back to the lowest-cost plan in the pruned tree; the deviant agent’s default violation penalty is 0.067 m.Source: §4.2 · pp. 4–5
• The oracle uses the same laws, planner, and scenarios as the social agent but replaces probe readings with simulation ground truth.This comparison isolates capability limits imposed by the simulated environment.Source: §4.2 · p. 5

• The planner receives RGB and proprioception, predicts cube position and sign color from latent state, and uses three-step pushes; cube rotation is locked.The cube’s pose is withheld from the planner and must be inferred from pixels by probes.Source: §5 · p. 5
• The Franka arm pushes a 0.09m cube across a 3×3 grid in eight opposite-cell tasks, each run over 50 variations (400 samples per agent).The center cell is illegal and never a start or goal.Source: §5 · p. 5
• The lawset makes cell 4 illegal by default, uses a yellow-cell checkpoint to unlock center access, and freezes the cube after a red-sign violation.Reaching the checkpoint flips the sign green after a clean trajectory or red if the cube previously passed through the center.Source: §5 · p. 5
• DDL supports R9’s permission conflict, in which an action is permitted both to happen and not to happen.Source: §5 · p. 5

• The compact lawset covers an earned exception to a defeasible prohibition, a compensatory duty, a permission conflict, and runtime rule insertion.Source: Table 1 · p. 6
• Planning is pruned using probe-predicted in_cell, while sign changes use ground-truth occupies and visited; runtime R11 obliges return to the starting cell.Source: R Rule R Rule R Rule · p. 6
• Prediction error can place the cube incorrectly and create false legal atoms, while R2 admits multiple planning-constraint interpretations.The latter gap can yield drastically different allowances for robot behavior.Source: R Rule R Rule R Rule · p. 6
• The first experiment runs laws 1–10 with a yellow checkpoint signal at t = 1; the second inserts R11 at t = 3 after forcing the sign white.Source: R Rule R Rule R Rule · p. 6
• Task success counts trajectories reaching the goal by cube center, whereas law abidance counts trajectories with no part of the cube entering an illegal cell.Source: R Rule R Rule R Rule · p. 6

Q1: Ex Ante Legislation

• At δ = 0.04, social abidance rose from 64.7% to 88.7%; baseline social abidance was 5.9× the realistic agent’s.Success rose alongside abidance with the uncertainty cushion.Source: Figure 2 · p. 7, §6 · p. 7

Q3: Endogenous Changes

• After R11 was inserted, the contrary-to-duty rule returned 114 cubes to legal cells by the next step, and 329 episodes returned to the start cell without compromising success.R11 was inserted at t = 3.Source: Figure 3 · p. 7, §6 · p. 7

Q2: Exogenous Signals

• When the sign permitted entry, the social agent entered the center on 53.5% of decisions, versus 7.4% when forbidden—a 7.2× increase.The sign changes according to ground-truth trajectory history.Source: §6 · p. 7

Q4: Grounding Isomorphism Gap

• WM+probe error averaged 0.031m (23% cell width), while 99th-percentile error reached 0.190m (143% cell width), enough to ground a false legal atom.Figure 4 pools error per executed action across the three agents.Source: §6 · pp. 7–8, Figure 4 · p. 8
• False-positive R10 repairs applied to cubes never in the illegal cell in 53% of social and 40% of deviant repair obligations; δ cushioning cannot recover the 143% error tail.The optimal cushion size is left to future work.Source: §6 · p. 8

Q5: Ontological Isomorphism Gap

• R2 metric choices shifted scored abidance by up to 32.3% for social, 23.9% for realistic, and 23.2% for deviant agents.Table 2 rescored the same episodes under alternative metrics rather than rerunning them.Source: §6 · pp. 8–9, Table 2 · p. 9

Q6: Auditability and Runtime Cost

• The audit ledger records decisions and ground-truth enforcement; R2 was enforced 91.1% for social and 87.3% for deviant, while reasoning took 16.6ms versus 12.5s for planning.Engaged indicates a conclusion was in force in the ledger; enforced means it also held in ground truth.Source: §6 · p. 9, Table 3 · p. 9

7. Related Work

• Prior work used deontic logic for learning-agent penalties, compliant goals, action supervision, and planning; this work applies it to sample-based motion planners in simulation for the first time.Earlier work also amended norms at runtime symbolically and without perception.Source: §7 · pp. 9–10
• Unlike pre-generation search filtering, the stack adjudicates generated nodes because legislation depends on observations and predictions, while DDL represents exceptions, duties, and violations.The approach is related to ethical governors, safety shielding, and temporal-logic planning.Source: §7 · p. 10
• Probe-grounded propositions keep verdicts traceable to their atoms; learned symbolic vocabularies and embedded temporal logic instead learn or shift the representation.The paper fixes its vocabulary from the legal source text.Source: §7 · p. 10

8. Discussion and Future Work

• The grounding gap is an engineering problem for deployment robustness, while the ontological gap requires legal and engineering expertise to translate even brightline rules into constraints.The latter lies between the judicial reasoning legislators may not be expected to provide and the jurisprudential work engineers may not be responsible for.Source: §8 · p. 10
• Future governance work needs better uncertainty prediction for perception-derived facts, including modeling tail prediction errors and state-specific biases.Conformal prediction is offered as one possible direction.Source: §8 · p. 10
• Physical-robot validation is critical for detecting real-world failure modes; scaling also requires principled mapping from open-texture or abstract laws to constraints.The authors suggest a Sim-to-Real Isomorphism gap likely exists.Source: §8 · p. 10

9. Conclusion

• The system implements robot governance tailored to world-model-based planning, using latent prediction space for ex ante governance.Source: §9 · p. 10
• Deployment-ready governance faces both grounding and ontological isomorphism gaps, with abidance varying widely across metric interpretations of an ostensibly brightline rule.Source: §9 · p. 10
• The conclusion argues that closing these gaps can help robot law align behavior with norms society sets, as legislatures intend.Source: §9 · p. 10