Daily Summary for 2026-10-23

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: It's the twenty-third of October, twenty twenty-six, and this is the day's research.

Dev: 30 new papers came out today.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We'll take the day in one pass, then pull out the papers we're staying with.

Dev: Start us off with SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Model.

The summary: Rosa: Welcome to the twenty-third of October, twenty twenty six. This is a look at making embodied intelligence safer by improving reinforcement learning for vision language action models.

Dev: SafeDojo proposes a model-based safe reinforcement learning framework using world-model imagination on top of an interactive video world model.

Rosa: This approach is significant because existing safe methods either demand costly real-world exploration or rely on manually designed safety functions, neither of which scales well for vision language action models in open environments.

Dev: The core mechanism involves decoupling task execution from trajectory safety using a ResNet success classifier and a lightweight safety head.

Rosa: This head predicts per-step contact probabilities from the current latent context and proposed action chunk, all balanced through Lagrangian-constrained GRPO.

Dev: This framework showed promising results, improving the average safe-success rate by 8.25 percentage points on SafeLIBERO Level I and remaining competitive with strong model-free baselines in zero-shot Level II settings.

Rosa: mu zero introduces a scalable world model based on 3D traces to enable robot learning without relying on embodiment-specific action labels.

Dev: This model forecasts smooth 3D trajectories for salient interaction points, and its TraceExtract system pretrains it by combining a vision-language backbone with a modular trace expert.

Rosa: For complex manipulation tasks, MiTaS explores fusing multiple heterogeneous tactile sensors operating at different temporal resolutions to solve contact-rich manipulation problems.

Dev: This framework uses modality-specific convolutional stems and transformer-based fusion to integrate information from RGB cameras, a GelSight Mini sensor, and an Evetac event-based sensor.

Rosa: MiTaS achieves an average success rate of 80% on five contact-rich manipulation tasks.

Dev: Finally, ASPIRE explores agentic skill programming through iterative robot exploration using a code-as-policy paradigm.

Rosa: ASPIRE autonomously writes and refines robot control programs in a closed-loop execution engine to diagnose failures and distill validated fixes into a reusable skill library.

Dev: ASPIRE has shown significant gains, surpassing prior methods by up to 77% on manipulation tasks under perturbation and achieving zero-shot generalization on representative unseen long-horizon tasks.

Rosa: The most pressing work here is about creating reliable forecasting tools for prosumer flexibility in ancillary service markets because without them, demand response aggregators cannot accurately bid into frequency controlled services where reliability standards like the P90 are very strict.

Dev: This paper proposes a scalable uncertainty quantification framework that combines Monte Carlo dropout with conformal prediction to give calibrated prediction intervals for aggregated prosumer flexibility.

Rosa: Conventional methods often fail; specifically, standalone Monte Carlo dropout methods systematically overestimate available flexibility and violate P90 compliance.

Dev: The proposed MCD-CP framework, however, achieves reliable coverage with controlled conservatism when embedded in the aggregator bidding model, substantially reducing overbidding risk and reaching up to seventy percent of perfect-information profit while meeting regulatory constraints.

Rosa: The work on action-conditioned world models addresses training autonomous tasks where demonstrations are hard to get.

Dev: It uses a latent conditional diffusion world model to predict future observations then feeds into a goal-conditioned temporal transformer.

Taro: Another relevance is how data center growth impacts grid resilience under disruptive events.

Rosa: This paper evaluates unserved energy when aggregated data center demand is added to an IEEE 30-bus system.

Dev: The results show temporally concentrated demand can amplify resilience impacts significantly even without increasing total consumption.

Taro: There is research into decentralized discrete-event systems where comparing local decision fusion rules is difficult.

Rosa: This paper extracts a method based on morphisms between graphs to provide a uniform approach for comparing architectures.

Dev: The most significant development involves coordinating traffic light systems and connected automated vehicles in mixed-traffic intersections.

Taro: Viet-Anh Le Andreas A Malikopoulos develops a distributed solution framework using a mixed-integer quadratic program for optimization.

Rosa: This approach uses proximal ADMM to solve the convex relaxation of the MIQP while employing a finite-grid tightening rule.

Dev: This is crucial for ensuring real-world feasibility when integrating autonomous vehicles into existing infrastructure.

Taro: This optimization builds on control theory by addressing stability in systems with limited regions of attraction.

Rosa: Lorenzo Fici Dalim Wahby Alvaro Detailleur Matthieu Barreau Guillaume Ducard's study investigates region of attraction estimation for linear quadratic regulators.

Dev: It uses both analytical bounds and Monte Carlo estimation to evaluate controller performance beyond conservative guarantees.

Taro: Research focuses on improving robotic decision-making by refining how models handle real-world discrepancies.

Rosa: Tuo An et al.'s Feedback World Model enables precise guidance of diffusion policy using an inference-time mechanism.

Dev: This closes the loop between prediction and observation by maintaining a lightweight feedback state to correct future predictions.

Taro: It has shown substantial improvements in prediction accuracy under distribution shift compared to static world modeling.

Rosa: Efforts are being made to make robotic learning more structured and interpretable for complex tasks.

Dev: Yiyuan Pan et al.'s Emergent Neural Automaton Policies introduces a framework for a bi-level neuro-symbolic policy.

Taro: This framework emerges from visuomotor demonstrations by first inferring a Mealy state machine to capture latent task modes.

Rosa: This structured approach shows promise in achieving high sample efficiency for long-horizon tasks.

Rosa: WARPED synthesizes wrist views to train policies with less data.

Dev: That reduces data collection time significantly, which is useful for training robots.

Taro: Safe-Night VLA integrates thermal perception into vision-language models for unseen areas.

Rosa: It uses control barrier functions during execution to enforce safety constraints reliably.

Dev: So foundation models can leverage non-visible physical modalities for robust manipulation then.

Taro: HEHA breaks complex Hamiltonian path problems into global and local planning stages effectively.

Rosa: Its routing algorithm RCH finds solutions forty-seven percent faster in makespan compared to existing methods.

Dev: AFFORD2ACT shows sparse keypoints can still retain strong manipulation capability across tasks.

Taro: Using learned gating with uniform weights caused success rates to drop significantly, however.

Rosa: AgriVLN introduces an STL module to break long language instructions into simple subtasks for navigation.

Dev: This improves performance on short commands while slightly increasing error on longer ones compared to existing methods.

Taro: Stochastic Neural Control Barrier Functions proposes frameworks synthesizing functions under different activation types smoothly.

Rosa: This builds upon the need for robust control, similar to DEFT's real-time branching using physics and learning.

Dev: The Cheap Hours Don't Stay Cheap shows how data center mobility alters electricity prices and grid resources.

Taro: Centralized dynamic state estimation distinguishes between cyberattacks and physical faults in power systems effectively.

Rosa: Evaluating Grid Resilience assesses how aggregated data center demand impacts system resilience under contingency events.

Dev: SafeDojo is a model-based safe reinforcement learning framework improving policies through world-model imagination.

Taro: Multi-Resolution Tactile Imitation Learning leverages multiple tactile sensors at different resolutions for contact tasks.

Rosa: ASPIRE is a continual learning system that writes and refines robot control programs by compounding experience into skills.

Dev: Inference-Time Robot Behavior Steering reconfigures task structure at test time based on user preferences without retraining.

Taro: A Uniform Approach to Compare Architectures in Decentralized Discrete-Event Systems provides a uniform comparison method for decentralized systems.

Rosa: OSCAR learns how long a robot should wait before rerouting around temporary obstacles for graph navigation.

Dev: Hilti-Trimble-Oxford Dataset combines 360 cameras and IMUs with LiDAR for robust visual simultaneous localization and mapping.

Taro: A Uniform Approach to Compare Architectures in Decentralized Discrete-Event Systems is worth hearing for comparing architectures.

Rosa: SafeDojo is a model-based safe reinforcement learning framework improving policies through world-model imagination.

Dev: Multi-Resolution Tactile Imitation Learning leverages multiple tactile sensors at different resolutions for contact tasks.

Taro: ASPIRE is a continual learning system that writes and refines robot control programs by compounding experience into skills.

Rosa: Inference-Time Robot Behavior Steering reconfigures task structure at test time based on user preferences without retraining.

Dev: OSCAR learns how long a robot should wait before rerouting around temporary obstacles for graph navigation.

Taro: Hilti-Trimble-Oxford Dataset combines 360 cameras and IMUs with LiDAR for robust visual simultaneous localization and mapping.

Rosa: A Uniform Approach to Compare Architectures in Decentralized Discrete-Event Systems provides a uniform comparison method for decentralized systems.

Dev: SafeDojo is a model-based safe reinforcement learning framework improving policies through world-model imagination.

Taro: Multi-Resolution Tactile Imitation Learning leverages multiple tactile sensors at different resolutions for contact tasks.

Rosa: ASPIRE is a continual learning system that writes and refines robot control programs by compounding experience into skills.

Dev: Inference-Time Robot Behavior Steering reconfigures task structure at test time based on user preferences without retraining.

Taro: OSCAR learns how long a robot should wait before rerouting around temporary obstacles for graph navigation.

Rosa: Hilti-Trimble-Oxford Dataset combines 360 cameras and IMUs with LiDAR for robust visual simultaneous localization and mapping.

Dev: A Uniform Approach to Compare Architectures in Decentralized Discrete-Event Systems provides a uniform comparison method for decentralized systems.

Lucky paper: 2606.20698: Rosa: So we're looking at SafeDojo today, which is about making embodied intelligence safer by improving reinforcement learning for vision language action models through world model imagination.

Dev: It tackles the problem of learning collision-free safe VLA policies in obstacle-rich environments where existing methods either require costly realworld exploration or depend on hand-crafted safety functions.

Rosa: SafeDojo extends standard VLA MDPs to Constrained MDPs by introducing a safety cost function c: S times A to zero one, requiring the policy pi theta to maximize J r (pi theta) subject to J c (pi theta) is less than or equal to d.

Dev: The paper proposes three key innovations: Chunk-wise Interactive World Model Rollouts, Decoupled Task Reward and Safety Cost Estimation, and Lagrangian-based Safety Constrained GRPO.

Rosa: For the world model rollouts, SafeDojo performs model-in-the-loop rollouts inside an interactive world model p phi which acts as an action-conditioned video predictor generating future visual observations and latent dynamics.

Dev: They use a recursive rollout formula for this: (hat oi k plus one z hat i k plus one) equals p phi (hat oi k z hat i k ai k l) for all k from zero up to K minus one.

Rosa: To make those rollouts robust against train-test mismatch caused by exposure error in long-horizon rollouts, they adopt a static-video augmentation to occasionally freeze the conditioning frames.

Dev: Then they decouple task reward and safety cost estimation by assigning dense feedback at every single control step, aggregating those into trajectory level pairs using a topM aggregation where M equals sixteen.

Rosa: The task reward branch ftask is instantiated with a ResNet success classifier finetuned on rollout frames, while the safety cost branch fsafe is a compact convolutional latent encoder followed by an action-conditioned MLP head that outputs per-step obstacle contact probability.

Dev: They then optimize these using a Lagrangian-based safety constrained GRPO objective, computing group relative advantages as A r i equals r i minus mu G of r over sigma G of r plus epsilon.

Rosa: The safety constraint is enforced by combining these advantages using a non-negative Lagrange multiplier eta to compute the safe advantage A safe i equals A r i minus eta A c i.

Dev: The multiplier eta is updated according to the batch level constraint violation: eta equals eta plus alpha eta over (one over G times G i equals one c i greater than d else eta decreases).

Rosa: The final objective is LSafe-GRPO, which minimizes the expected return of the safe advantage clipped between rho i t theta, one minus epsilon, and one plus epsilon.

Dev: On SafeLIBERO Level I, SafeDojo achieved an eight point two five percentage-point improvement over the strongest baseline on Level I in the average safe-success rate.

Rosa: In realworld Franka deployment, SafeDojo further shows the best average task and safe-success rates across five tasks.

Dev: They also did an ablation study confirming that both components are critical for performance; removing the safety head reduces SSR to thirty-eight point zero percent.

Rosa: And removing the task progress reward branch collapses TSR and SSR to three point zero percent and increases ETS to two hundred ninety-eight point three, which indicates task-reward feedback is essential for goal-directed policy improvement.

Taro: From a world model perspective, how does this imagined trajectory feedback help bridge the gap between simulation and realworld execution?

Rosa: It lets the system learn safe actions through world model imagination instead of relying on static safety rules or risky real-world exploration, which is a big deal for scaling.

Dev: SafeDojo establishes world model-based safe RL as a scalable and generalizable path toward safe embodied intelligence by learning both task progress and safety signals from imagined trajectories.

Taro: The limitation you mentioned about the safety head being trained on binary collision labels might not capture finer grained safety requirements like force limits or proximity constraints.

Rosa: That is a known limitation, but the paper acknowledges that the world model introduces prediction errors that accumulate over longer rollout horizons, potentially degrading both task progress and safety evaluation.

Dev: The current evaluation focuses on tabletop manipulation with static obstacles, and deploying SafeDojo in more dynamic environments remains an open challenge.

Taro: What about inference time? Does this framework offer any hard safety guarantees when the robot is actually running?

Rosa: Currently, SafeDojo operates as a training-time safety mechanism and does not provide hard safety guarantees at inference time.

Dev: Future work includes integrating learned safety signals with inference-time safeguards such as control barrier functions to address that.

Taro: It also moves the focus from just learning a policy to inferring symbolic structure from visuomotor demonstrations using Emergent Neural Automaton Policies.

Rosa: That framework shows promise in achieving high sample efficiency for long-horizon tasks by inferring a Mealy state machine to capture latent task modes first.

Dev: And they use WARPED to synthesize realistic wrist view observations from monocular RGB data, requiring significantly less data collection time than previous methods.

Taro: It’s interesting how they are pushing both the safety mechanism and the structural learning aspects simultaneously here.

Lucky paper: 2607.00272: Rosa: Okay, we are looking at ASPIRE now, which is about agentic skill discovery for robotics.

Dev: This system is a continual learning system that autonomously writes and refines robot control programs in a code-as-policy paradigm while compounding experience into a reusable skill library.

Rosa: It operates in an open-ended learning loop with three key components: the robot execution engine, the skill library, and an evolutionary search procedure.

Dev: The robot execution engine replaces coarse rollout-level feedback with per-primitive execution traces for each primitive call.

Rosa: For every primitive call, the trace stores the invoked API, inputs and outputs, return status, and multimodal evidence like RGB keyframes or grasp candidates.

Dev: These rich multimodal traces let the agent selectively inspect salient primitive logs to localize failures.

Rosa: For instance, in a navigation task where candidate navigation goals fall inside the collision-avoidance buffer due to table constraints, the trace localizes it to "repeated PLANNING ERRORs."

Dev: The agent then patches the program with a multi-angle approach routine, and this validated repair is admitted as a reusable Multi-Angle Approach skill.

Rosa: The skill library maintains a growing library that distills validated fixes into reusable knowledge retrievable as in-context guidance for future tasks.

Dev: Each skill is stored as compact in-context guidance including the failure signature, when-to-apply condition, repair strategy, and a representative code sketch.

Rosa: The coordinator audits these findings and promotes only reusable repairs that have passed debug validation into the shared library.

Dev: ASPIRE employs an evolutionary search procedure to generate diverse task sequences and control programs exploring beyond single-trajectory self-improvement.

Rosa: In each round, based on the skill library, the coding agent proposes a population of K candidate programs conditioned on the top-performing previous programs and failure traces from previous evaluations.

Dev: The search target is the robot program itself, and candidates are selected through closed-loop execution to explore distinct strategies rather than repeatedly refining the same solution.

Rosa: This procedure allows ASPIRE to scale its performance with experience as "the more tasks Aspire sees, the larger its skill library grows."

Dev: The empirical results show strong self-improvement across diverse short and long horizon robot benchmarks.

Rosa: Against prior coding agents, ASPIRE improves success rate by up to seventy-seven points on LIBERO-Pro under perturbation suites.

Dev: It also improves success rate by up to seventy-two points on Robosuite contact-rich manipulation tasks and by up to thirty-two percent on long-horizon household tasks.

Rosa: Based on the skill library accumulated on LIBERO-ninety ASPIRE transfers zero-shot to LIBEROPro Long with thirty-one percent success.

Dev: This substantially outperforms the four percent success rate of prior methods despite their heavy reliance on test-time reasoning and retries.

Taro: The limitations include not yet being a fully autonomous real-world lifelong learner because it needs robust success detection and safe reset in the physical world.

Rosa: It also relies on a frozen frontier LLM, Claude Opus four point six with a 1M-token context window, to interpret multimodal traces and write program repairs.

Dev: Finally, the debug and evolutionary-search loop is compute-intensive because it consumes many LLM calls and simulator or robot rollouts per task.

Taro: The agent pipeline involves a coordinator managing the shared skill library and dispatching actor coding agents to individual tasks in a parallel manner.

Rosa: The subagent's role is to diagnose failed robot trials for one task, write a generalizable repair program, validate it according to the protocol, and report concise results.

Dev: The coordinator then promotes generalizable patterns into the shared skill library after each completion.

Lucky paper: 2606.00990: Rosa: Okay, we're moving on to OSCAR: Obstacle Survival Curves for Adaptive Robot Navigation. This paper looks at how a robot learns to adaptively decide when to wait versus reroute when a temporary obstacle blocks an edge in graph-based navigation.

Dev: So the core idea is that the robot learns class-conditioned residual clearance-time distributions from online experience.

Rosa: Exactly, it assumes we have obstacle class labels available at the time of encounter, and it fits a non-parametric Kaplan–Meier survival model for each obstacle class.

Dev: It supports learning from both clearance observations and right-censored observations when the robot reroutes before seeing clearance.

Taro: The decision-making module needs three things: a survival model predicting how long each obstacle class tends to remain, an edge cost converting those predictions into expected navigation delay, and a local wait-or-reroute rule choosing the patience threshold Wstar.

Rosa: The robot selects this patience threshold Wstar by minimizing the expected time-to-goal under that set of survival estimates.

Dev: They define time-dependent edge costs as bc(e, t) equals w(e) plus delta(be, t), which is the nominal traversal time plus expected extra delay from obstacles.

Taro: For edges with a remembered obstacle state M(e), the expected additional wait is calculated using qe of t times mu old e of t.

Rosa: This results in a delay that's a mixture: with probability qe of t, the old obstacle is still present causing residual delay, and otherwise the edge is treated like an ordinary one.

Dev: When a blocked edge is reached, it chooses a patience threshold W means wait up to W seconds; if it clears before then, it reroutes with the edge available.

Taro: If the obstacle doesn't clear within that time, it replans with the blocked edge forbidden.

Rosa: They choose Wstar by minimizing J(W), which is the expected remaining time-to-goal under threshold W, and they search for Wstar in zero Wmax.

Dev: The search for this patience threshold reduces to evaluating Jb(W) on a finite candidate set of survival event times and the horizon endpoints.

Taro: In simulation, the learned policy’s time-to-goal converges to within one percent of an oracle with access to ground truth clearance distributions after fewer than twenty observations per obstacle class.

Rosa: Real-world deployment confirms that the policy improves online, adapting its patience thresholds across fifty navigation episodes.

Dev: OSCAR achieves a mean time-to-goal of one hundred sixty point seven seconds and shows an increase in success rate when reroutes are counted alongside waiting time for blocked edges.

Taro: In realworld experiments, the time-to-goal decreases over repeated episodes and converges to approximately two hundred seconds, outperforming the always-wait baseline.

Rosa: One limitation is that the current wait-or-reroute decision estimates the cost of committing to a different path, not accounting for encountering another obstacle later.

Dev: Another issue is that outcome-adaptive censoring can violate standard independence assumptions when computing survival curves.

Taro: Finally, using a single learned blockage probability for each obstacle type across all edges might not account for spatial layout differences between edges.

Rosa: The conclusion is that by learning class-conditioned obstacle-clearance models from online observations, the robot can adaptively decide how long to wait before rerouting.

Dev: More broadly, these results suggest that modeling obstacle persistence is a useful direction for long-horizon navigation in robotics.

Lucky paper: 2607.06464: Rosa: We're looking at the Hilti-Trimble-Oxford Dataset today, which is a construction focused visual inertial state estimation benchmark with prior 2D maps.

Dev: This dataset has thirty visual-inertial sequences recorded across seven floors over an eight month period of the construction project.

Rosa: Those sequences capture variable lighting conditions, moving workers, fast motions, and repetitive structures.

Dev: The data includes thirty visual-inertial sequences with synchronized IMU measurements captured over a span of eight months at the same construction site.

Rosa: They are paired with corresponding 2D floor plans which are binary images where structural elements are encoded as one and free space as zero.

Dev: Ground-truth trajectories were collected using a high quality LiDAR-inertial SLAM system rigidly attached to the three hundred sixty camera.

Rosa: That ground truth system consists of a Hesai XT32M2X LiDAR sensor and a high quality, factory calibrated, tactical grade IMU with hardware synchronization between the two components.

Dev: Temporal alignment is achieved by estimating a linear time-mapping function between the LiDAR and camera clocks by maximizing the crosscorrelation between their corresponding inertial motion signals.

Rosa: Geometric optimization involves an online MC2SLAM algorithm followed by a joint optimization in a large scale non linear optimization step.

Dev: The evaluation protocol supports two categories of challenges: one is in an arbitrary reference frame for SLAM and the other is in the floor plan’s reference frame for localization.

Rosa: Submissions must achieve at least ninety-nine percent coverage of the evaluated ground-truth poses to be scored, with an exponentially decaying position score computed over the trajectory.

Dev: The final score is calculated as the sum of the trajectory scores over all evaluated runs.

Taro: The Hilti x Trimble Challenge two thousand twenty-six attracted substantially higher participation in SLAM with sixty-two teams compared to twenty-two in floor-plan-referenced localization, reflecting the broader maturity of SLAM methods.

Rosa: The winning SLAM submission achieved a total score of two thousand four hundred ten point zero points.

Dev: The top team in the Localization category achieved an average error below twenty-four cm.

Taro: The top SLAM solutions generally rely on visual-inertial odometry usually using both cameras and the IMU and followed by a subsequent optimization step BA.

Rosa: Furthermore, the localization track results suggest that semantic information extraction plays an important role in map alignment.

Dev: A primary issue noted is that the bitmap floor plans were derived from a single version of the architectural floor plans.

Taro: That means the Challenge required localization to a similar but different scene with only some of the features usable for matching.

Rosa: Additionally, a few deviations or errors in the floor plan were identified when comparing them to our LiDARbased ground-truth point clouds.

Lucky paper: 2210.16511: Rosa: So we're looking at A Uniform Approach to Compare Architectures in Decentralized Discrete-Event Systems.

Dev: It starts by formally defining decentralized control problems, which they call Problem two point one, which involves constructing local decision functions and fusion rules meeting certain conditions regarding languages K and L.

Rosa: Then Problem two point two introduces the Decentralized Observation Problem, which requires constructing local decision functions for agents i in N and a fusion rule f such that for all s in L, if s is in K then f of f1 P1(s),..., fn Pn(s) equals one but if s is in L minus K then it equals zero.

Dev: Theorem two point three establishes an equivalence between the control problem C and a set of observation problems O for each sigma c i.

Rosa: The comparison of architectures is then stated as determining the relationship between the sets Of and Og; if all problems solvable by f are solvable by g, meaning If Of subseteq Og, then g is at least as permissive as f.

Dev: The paper aims to derive a uniform approach for easy and direct comparison of decentralized architectures without needing to pass indirectly through characterizations and counterexamples.

Rosa: They describe a decentralized problem as an observation graph and a solution based on a fusion rule f as a decision graph, expressing the problem as finding "a way of folding the observation graph into the decision graph" formally in terms of "graph morphism".

Dev: The solvability condition is then thought of as determining if the decision graph has the capacity to embed the observation graph.

Rosa: They introduce two core concepts: an Observation Graph defined by a symmetric relation N on L, where s is related to t if and only if the tuples P1(s),..., Pn(s) and P1(t),..., Pn(t) differ by exactly the components indexed by N.

Dev: This forms an undirected graph (L, N) called the observation graph, where nodes are strings in L and edges are colored by subsets of N.

Rosa: The Decision Graph is defined for a fusion rule f over a set D of combinations of local decisions d1,..., dn using symmetric relations on D, where (d1,...,dn) is related to (d'one...,...) when the two tuples differ by exactly the components indexed by N.

Dev: This forms an undirected graph (D, N) called the decision graph.

Rosa: The relationship between these two graphs is formalized through a mapping g = <fiPi>, which is a type of "graph morphism".

Dev: This mapping must satisfy GM-one Node-Colour Preserving, meaning the mapping preserves node colouring and achieves the desired fused decision.

Rosa: And GM-two Edge-Colour Intensive, meaning N superset of N prime i.e., only agents observing s and t differently can issue different decisions for the two strings.

Dev: This property captures feasibility in terms of a solution existing for the observation problem.

Rosa: Theorem three point five states that an Observation Problem O has a solution if and only if there exists a morphism from the observation graph to the decision graph satisfying these GM conditions.

Dev: Theorem four point one provides the main result for comparing architectures directly: fusion rule f prime is more permissive than f, meaning If O(f, L, K, Pi) is in Of, then O(f', L, K, Pi) is in Of prime if and only if there is a graph morphism from D to D' satisfying the graph morphism conditions GM.

Rosa: This result relies on Lemma four point two which states that for every decision graph (D, N), there is an isomorphic observation graph where the isomorphism preserves node- and edge-colouring.

Dev: The proof shows that if a morphism g from D to D' satisfies GM, it implies that all problems solvable with f are also solvable with f prime.

Rosa: Conversely, if all observation problems solvable with f are solvable with f prime, then there must exist a morphism from D to D' satisfying GM.

Dev: They illustrate this comparison through examples; for instance, Example four point three shows how the C andP and D andA architecture is weaker than the C andP architecture by giving a decision graph morphism that sends all nodes in the dashed box to (one one).

Rosa: Example four point five demonstrates incomparability between the conjunctive architecture f = lambda and the disjunctive architecture f = sigma by showing that no morphism exists in either direction between their respective decision graphs.

Dev: In summary, this work proposes two tools—observation graphs and decision graphs—to study decentralized observation problems, providing a systematic approach to directly compare decentralized architectures through graph morphisms.