Active Inference and Behavior Trees for Reactive Action Planning and Execution in Robotics

May combine two claims.

ROSPlan is not designed for fast reaction to dynamic environmental changes caused by external events, and it would not work when the initial state is only partially observable.

“ROSPlan leverages automated planning with PDDL2.1, but it is not designed for fast reaction in case of dynamic changes in the environment caused by external events, and would not work in case of a partially observable initial state.”

• cites [41] sourceDisputed

  The Case study (section 4; page not provided) says the AUV began without knowing the tank size, obstacles, or valve-panel location. The Outcome section reports that discovering the panel canceled the current action and prompted replanning, contrary to the claim that ROSPlan would not work with an incompletely known initial state or adapt to environmental changes. The paper does not establish a measured speed of reaction. (LLM, conf=0.86)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

May combine two claims.

The method removes the need to detect conflicts explicitly and reorder the BT, unlike earlier dynamic BT expansion.

“With our method, there is no need to explicitly detect a conflict and reordering a BT then, as was done in past work on the dynamic expansion of BTs [4].”

• cites [4] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Active inference is a general framework covering perception, action, decision-making, and learning in biological or artificial agents.

“Active inference provides a unifying theory for perception, action, decision-making, and learning in biological or artificial agents [33].”

• cites [33] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

RLDS gives a more concise problem description than behavior trees alone, but its online reactivity is limited to offline-planned scenarios.

“RLDS results in a more concise problem description with respect to using BTs only, but online reactivity is again limited to the scenarios planned offline.”

• cites [6] sourcePartly supported

  The Introduction and Related Work describe RLDS task descriptions as simpler than specifying a BT’s internal structure. The Execution section says plan construction occurs offline and policies are selected in near real time. But the paper does not say reactivity is limited to scenarios anticipated offline; the Introduction and Real-World Experiments describe adaptation to unforeseen interference. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Discrete-state active inference is a promising way to address the exploration-exploitation dilemma and gives agents continuous deliberation.

“Active inference on discrete state spaces [33] is a promising approach to solving the exploitationexploration dilemma and empowers agents with continuous deliberation capabilities.”

• cites [33] sourcePartly supported

  In “Planning and decision-making” (page unavailable), the paper says expected free energy balances exploitative and explorative courses of action, supporting the exploration–exploitation part of the claim. It describes agents evaluating policies and planning, but the passages do not establish that discrete-state active inference gives agents continuous deliberation capabilities. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

May combine two claims.

The blended reactive task and action planner from [4] handles unanticipated scenarios but not partially observable environments or uncertain action outcomes.

“The solution can compensate for unanticipated scenarios, but cannot handle partially observable environments and uncertain action outcomes.”

• cites [4] sourcePartly supported

  The Introduction (p. 1) describes the planner reacting when an object slips from the gripper or an obstacle unexpectedly blocks its path. The Problem Formulation (p. 4) also allows it to try another action after one fails. Our copy does not establish that the approach cannot handle partially observable environments or uncertain action outcomes; it describes some responses to unexpected outcomes instead. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

May combine two claims.

Earlier active inference studies took actions to be always executable and conflict-free, but the authors say this is false in realistic settings.

“Past work on active inference, such as [23], was based on the assumption that actions were always executable and nonconflicting, but these do not hold in more realistic scenarios.”

• cites [23] sourceCould not verify

  Sections 2 and 11 discuss action selection and earlier active-inference work, but do not say that earlier studies assumed actions were always executable and nonconflicting. Section 6 describes what happens when a movement meets a maze boundary; it does not establish the claimed assumption or discuss conflicting actions. Searches for executable, feasible, and conflicting actions found no decisive passage. Our copy provides no page numbers for these passages. (LLM, conf=0.74)

  Experimental — accuracy not yet measured

May combine two claims.

The vector E can encode common sense and habits and can also be used to adapt plans online, for instance after a runtime component failure.

“This vector can be used to encode common sense and habits [34], [35], but can also be used to adapt the plans online due for instance to runtime component failure.”

• cites [34] Smith, 2022, A step-by-step tutorial on active inference and its application to empirical data, J. Math. Psychol. sourcePartly supported

  Section 2.1 (Formal POMDP structure; page not provided) says the E vector encodes prior beliefs about policies and can model habits. Our copy does not establish that E encodes common sense. The sentence places the citations before the separate claim about online adaptation after component failure, so that claim is not attributed to this source. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [35] sourcePartly supported

  The Methods section (“An Incremental Primer on Active Inference”; page unavailable) identifies Eπ as a baseline prior over policies that can be thought of as habits. It does not establish that E specifically encodes common sense. The sentence places the citations before the separate claim about adapting plans after runtime failure, so that claim is not treated as attributed to this source. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

May combine two claims.

To change the attraction region, one must either replan or expand the BT dynamically.

“Alternatively, adapting the region of attraction requires either replanning [6] or dynamic BT expansion [4].”

• cites [6] sourcePartly supported

  The Simulation Experiments section describes replanning when execution conditions fail and says this lets the robot adapt. The paper does not state that changing a region of attraction requires replanning; the dynamic BT expansion alternative is attributed to a different citation in the citing sentence. No page number is available in our copy. (LLM, conf=0.77)

  Experimental — accuracy not yet measured
• cites [4] sourcePartly supported

  The Problem Formulation (p. 4) says the BT can be expanded during runtime when a previously assumed condition fails, supporting the dynamic-expansion part attributed to this paper. The copy does not say that this adapts a region of attraction or that expansion is required to do so. The separate replanning citation [6] is not assessed here. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Online decision-making without offline computations can achieve performance similar to active inference.

“Second, one could perform online decision-making without offline computations and achieve similar performance, see [44] and [45].”

• cites [44] sourceNot checked — no full text of the source
• cites [45] sourcePartly supported

  Section 2.2 (page unavailable) describes online planning as an alternative to offline policy computation, and Section 5.1 reports that DESPOT matched some POMDP planners in performance. The paper does not compare it with active inference. Moreover, Section 5.1 says DESPOT's parameter was selected through offline training, so its reported performance does not establish the claim's ‘without offline computations’ qualifier. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

Active inference can unify abstract decision-making with guarantees, adaptive and fault-tolerant torque control, and state estimation and learning.

“Active inference can unify 1) abstract decision-making with guarantees, as in this article, 2) adaptive [12] and fault tolerant [10], [11], [46] torque control, as well as 3) state estimation and learning.”

• cites [12] sourceSupported
• cites [10] sourceNot checked — no full text of the source
• cites [11] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

In Example 6, the amount of knowledge needed to code a classical BT and the active inference version is the same.

“Note that the amount of knowledge (i.e., number of states and actions) that is necessary to code a classical BT or our active inference version in Example 6 is the same.”

May combine two claims.

The results showed better runtime adaptability with far fewer hand-coded nodes than classical behavior trees.

“The results showed improved runtime adaptability with a fraction of the hand-coded nodes compared to classical BTs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Scenario 2 reused the Scenario 1 BT completely, changing only the target object and locations.

“Note that the BT designed for Scenario 1 was entirely reused in Scenario 2, with the only adaptation of the desired object and locations in the BT.”

May combine two claims.

The proposed method yields continual online planning and hierarchical deliberation.

“This results in continual online planning and hierarchical deliberation.”

• Also stated in: § B. Contributions · plain

  “to obtain more reactive actors with hierarchical deliberation and continual online planning capabilities.”

May combine two claims.

The approach abstracts fallback behavior by planning in state space rather than action space; instead of programming moveTo(goal), one sets a prior preference over isAt(goal).

“However, we abstract the fallback by planning in the state space and not in the action space.Instead of programming the action moveTo(goal) we only set a prior preference over the state isAt(goal) since the important information is retained in the state to be achieved rather than in the sequence of actions to do so.”

May combine two claims.

The result is conflict resolution and plan continuation that arise naturally, without special recovery machinery.

“This leads to a natural conflict resolution and plan resuming without ad hoc recovery mechanisms.”

May combine two claims.

Highly reactive behavior trees tend to be large and complex, and adding new skills requires revising a large tree.

“This means that highly reactive BTs are usually big and complex and that adding new robotic skills would require revising a large tree.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

GOAP needs scenario-specific hand-designed heuristics and is computationally expensive for long-term tasks.

“However, GOAP requires hand-designed heuristics that are scenario-specific and it is computationally expensive for long-term tasks.”

• cites [26], [27]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

When no goal exists, all state preferences are zero, giving no reason to act toward another state.

“Note that if there is no goal, the preferences over states are set to zero everywhere so there is no incentive to act to achieve a different state.”

May combine two claims.

The authors introduce a way to include action preconditions and conflict resolution in active inference and a systematic way to provide prior preferences through behavior trees.

“We introduce a method to include action preconditions and conflict resolution in active inference, as well as a systematic way of providing prior preferences through BTs.”

• AssumptionRests on an assumption: The continuous world state and the robot's internal state can be abstracted into symbolic states and observations that a discrete active inference agent can use.

May combine two claims.

An agent using the method can follow an offline plan while adapting locally and making autonomous runtime decisions within safety constraints.

“By doing so, an agent can follow a predefined offline plan while still keeping the ability to locally adapt and take autonomous decisions at runtime, respecting safety constraints.”

• Also stated in: § D · plain

  “Such an actor is capable of following a task planned offline while still being able to take autonomous decisions at run-time to resolve unexpected situations.”

Classified as probably not a substantial claim. May combine two claims.

HTN requires designers to write and debug complex domain-specific recipes, and replanning may happen too often in highly dynamic situations.

“HTN requires the designer to write and debug potentially complex domain-specific recipes and, in very dynamic situations, replanning might occur too often.”

• cites [31] Erol, 1994, HTN planning: Complexity and expressivityNot checked yet

  Not checked in the free run.

May combine two claims.

This replacement allows offline plans to be simplified, opportunities to be exploited, and local unforeseen contingencies to be resolved intelligently.

“This will allow us to simplify our offline plans, exploit opportunities, and act intelligently to resolve local unforeseen contingencies.”

May combine two claims.

The hybrid scheme uses online action selection with active inference, removes the need for complex predefined fallback behaviors, and provides convergence guarantees.

“The proposed hybrid scheme leverages the advantages of online action selection with active inference and it removes the need for complex predefined fallback behaviors while providing convergence guarantees.”

May combine two claims.

Current discrete active inference solutions lack a systematic task-independent method for specifying prior preferences and do not consider action preconditions needed in real robotics.

“In general, current discrete active inference solutions lack a systematic and task-independent way of specifying prior preferences, which is fundamental to achieving a meaningful behavior, and they never consider action preconditions that are crucial in real-world robotics.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Active inference has shown promise in control engineering and robotics, including real-world low-level adaptive control experiments.

“Active inference is a neuroscientific theory that has recently shown its potential in control engineering and robotics [8], [9], [10], [11], particularly in real-world experiments for low-level adaptive control [12], [13].”

• cites [8], [9], [10]Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S4 (based).

Mobile manipulation in dynamic environments is difficult because offline-planned actions often fail.

“This is particularly useful for challenging problems, such as mobile manipulation in dynamic environments, where actions planned offline are prone to fail.”

May combine two claims.

Dynamic priors cut down on hard-coded fallbacks and let the BT shrink to few nodes.

“This reduces considerably the need for hard-coded fallbacks, allowing to compress the BT to a minimal number of nodes.”

Probably not a substantial claim: may not pass S4 (based).

If no available action leaves Lc(τ) in D(aN), the algorithm fails to meet the initial action's preconditions; the paper calls this a major failure for the BT to handle.

“Note that if L c (τ ) ∈ D(a N ) does not hold after sampling all available actions, it means that the algorithm is unable to find a set of actions that can satisfy the preconditions of the initially planned action.This situation is a major failure that needs to be handled by the overall BT.”

• AssumptionThe authors state this as an assumption: The action templates with pre- and postconditions provided to the agent are correct.
• AssumptionThe authors state this as an assumption: A given desired goal is achievable by at least one atomic action.

May combine two claims.

Robotic tasks can be described in terms of free-energy minimization, and action preconditions and conflict resolution for active inference were introduced by means of dynamic priors.

“We showed how robotic tasks can be described in terms of free-energy minimization, and we introduced action preconditions and conflict resolution for active inference by means of dynamic priors.”

May combine two claims.

Without knowledge of the full initial state, a ROSPlan-based solution would have to assume values for the unknown states and either plan for the worst case or fail during execution and replan.

“Without knowing the full state at the start, a solution with ROSPlan would require making assumptions on the value of the unknown states, and either planning for the worst-case scenario, which might not even be needed, or failing during execution and replan.”

May combine two claims.

The proposed solution provides local reactivity to unforeseen situations while keeping the initial plan intact.

“Our solution provides local reactivity to unforeseen situations while keeping the initial plan intact.”

May combine two claims.

Classical BTs usually fix attraction regions offline, which requires lengthy contingency analysis and is task-specific.

“When considering task planning and execution with classical BTs, often these regions of attraction are defined offline leading to a complex and extensive analysis of the possible contingencies that might eventually happen [3], and these are specific to each different task.”

• cites [3] sourceNot checked — no full text of the source

May combine two claims.

Behavior trees are advantageous because they dispatch and monitor action execution at runtime, which allows quick reaction to environmental changes that are not a consequence of the robot's own actions.

“In this context, BTs are advantageous because they are designed to dispatch and monitor actions execution at runtime.”

• Also stated in: § B. Why Choosing BTs · plain

  “This allows to quickly react to changes in the environment that are not necessarily a consequence of the robot's own actions.”

May combine two claims.

Active inference adaptation can be limited to task portions that need it, while preserving all BT properties.

“In fact, we are able to include adaptation through active inference only in the parts of the task that require it, keeping all the properties of BTs intact.”

May combine two claims.

The algorithm builds complex attraction regions online, reducing the need to preprogram all fallbacks in a BT that may fail in unforeseen edge cases.

“With our algorithm we can generate complex regions of attraction dynamically at runtime, alleviating the burden of programming every fallback beforehand in a BT, which is prone to fail in edge cases that have not been considered at design time.”

May combine two claims.

Rather than coding many fallbacks, Algorithm 2 gives the actor deliberation, letting it reason over the present state, goal, and action templates.

“Instead of extensively programming fallback behaviors, Algorithm 2 endows our actor with deliberation capabilities and allows the agent to reason about the current state of the environment, the desired state, and the available action templates.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Computing the model evidence exactly is a well-known and often intractable problem in Bayesian statistics.

“Computing the model evidence P (ō) exactly is a well known and often an intractable problem in the Bayesian statistics.”

Probably not a substantial claim: may not pass S4 (based).

Given Lc(τ) in D(a1), the algorithm builds a plan π = {a1,...,aN} whose domain of attraction D(π) follows the listed steps.

“If L c (τ ) ∈ D(a 1 ), Algorithm 2 generates a plan π = {a 1 , . . ., a N } with domain of attraction D(π) according to the steps as follows.”

• AssumptionThe authors state this as an assumption: The model parameters, such as likelihood and transition matrices, are known.
• AssumptionThe authors state this as an assumption: The action templates with pre- and postconditions provided to the agent are correct.
• AssumptionThe authors state this as an assumption: A given desired goal is achievable by at least one atomic action.

Probably not a substantial claim: may not pass S2 (checkable).

Active inference is a flexible and unified framework that connects different branches of control theory at different abstraction levels.

“We are particularly interested in active inference because it is a flexible and unified framework that connects different branches of control theory at different abstraction levels.”

May combine two claims.

The method keeps control of the robot's overall behavior and can force a routine when problems arise, using the BT structure.

“The proposed algorithm allows to retain control over the general behavior of the robot and to force a specific routine in case something goes wrong leveraging the structure of BTs.”

May combine two claims.

The hybrid approach drastically reduces the number of nodes planned offline in a behavior tree, handles partial observability of the initial state, and allows adaptation and safety guarantees to be added selectively in specific parts of the tree.

“In particular, it drastically reduces the number of necessary nodes planned offline in a BT, it can handle partial observability of the initial state, and it allows to selectively add adaptation and safety guarantees in specific parts of the tree.”

Probably not a substantial claim: may not pass S3 (load-bearing).

PDDL can be used to automatically generate plans whose execution is optimized through behavior trees.

“For instance, PDDL can be used to automatically generate plans for which their execution is optimized through BTs [42].”

• cites [42] Martn, 2021, Optimized execution of PDDL plans using behavior treesNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

BTs are reactive: each action or subbehavior fires in response to an event or condition at the current time step.

“BTs are purely reactive which means that every action, or subbehavior, is executed in response to an event or a condition determined at the current time step τ .”

May combine two claims.

The method is said to be more robust than classical BTs both for executing tasks and for handling noisy sensor readings.

“With our approach we improve robustness compared to classical BTs in two different ways: 1) in terms of task execution, and 2) against noisy sensory readings, which we elaborate in the following.”

May combine two claims.

Algorithm 2 obtains robustness by building an appropriate attraction region online through free-energy minimization, covering just enough to guide the present state to the goal while altering priors during execution.

“With Algorithm 2 we achieve robust behavior by dynamically generating a suitable region of attraction according to the minimization of free-energy.This region brings the current state toward the desired goal.We then cover only the necessary region in order to be able to steer the current state to the desired goal, changing prior preferences at run-time.”

• AssumptionThe authors state this as an assumption: The model parameters, such as likelihood and transition matrices, are known.

May combine two claims.

In that approach the whole behavior is determined at runtime from a goal condition, so safety guarantees can only be added for the whole task and not for selected parts of an online-derived tree.

“The whole behavior is indeed determined at runtime based on preconditions and effects of actions starting from a goal condition.”

• Also stated in: § APPENDIX E UPDATING PLAN DISTRIBUTION · plain

  “The addition of safety guarantees can only happen for the whole task, and not in selected parts of the tree derived online.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Specifying the likelihood and transition matrices that encode action pre- and postconditions for active inference is done once when defining a robot's available skills and is independent of the task to be solved.

“It is true that active inference requires specifying the likelihood and transition matrices encoding actions pre-and postconditions, but this has to be done only once while defining the available skills of a robot, and it is independent of the task to be solved.”

May combine two claims.

Adding new symbolic actions, or substantially changing the task while keeping the same skills, would require recomputing an offline policy.

“However, the addition of new symbolic actions (so new skills as transitions), or a substantial change in the task while using the same skills, would require recomputing the policy.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A theoretical analysis of the convergence and robustness of the algorithm was performed.

“We performed a theoretical analysis of the convergence and robustness of the algorithm”

May combine two claims.

Active inference makes use of the full probabilistic information on the states and does not require full knowledge of the symbolic initial state.

“However, by means of active inference, one can make use of the full probabilistic information on the states, and one does not require full knowledge of the symbolic initial state.”

Probably not a substantial claim: may not pass S4 (based).

Safety guarantees that override the online decision-making process can be easily added thanks to the properties of behavior trees.

“In addition, it is possible to easily add safety guarantees to override the online decision-making process thanks to the properties of BTs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The variational free-energy is an upper bound on surprise.

“We have just defined the free-energy as the upper bound of surprise F [Q(s, π)] ≥ -ln P (ō).”

• AssumptionThe authors state this as an assumption: The parameters associated with the task are known and introduce no uncertainty in the derivations.
• AssumptionThe authors state this as an assumption: Observations are conditionally independent of the model parameters given the state.
• AssumptionThe authors state this as an assumption: Under the Markov property, the next state and the current observations depend only on the current state.
• AssumptionThe authors state this as an assumption: The states and the model parameters are conditionally independent given the plan.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

When Lc(τ) lies in D(aN), the composed plan π, with attraction region D(π)=∪ D(ai), stabilizes at C; and if the actions are FTS, π is FTS.

“If L c (τ ) ∈ D(a N ), the sequential composition π with region of attraction D(π) = a i D(a i ) for i = 1, . .., N stabilizes the system at the given desired state C. If a i ∈ A are FTS, then π is FTS.”

• AssumptionThe authors state this as an assumption: The model parameters, such as likelihood and transition matrices, are known.
• AssumptionThe authors state this as an assumption: The action templates with pre- and postconditions provided to the agent are correct.
• AssumptionThe authors state this as an assumption: A given desired goal is achievable by at least one atomic action.

May combine two claims.

The proposed algorithm and its core idea are general and independent of the particular robot platform and task.

“The proposed algorithm and its core idea are general and independent of the particular robot platform and task.”

May combine two claims.

If Algorithm 2 finds an executable aN and forms π = {a1,...,aN}, the plan's attraction domain contains Lc(τ); it is asymptotically stable when D(a1) is contained in the union of Pc(ai) for i=2,...,N.

“When an executable action a N is found during task execution through Algorithm 2 such that π = {a 1 , . . ., a N }, the plan has a domain of attraction toward a given goal that includes the current state L c (τ ).If D(a 1 ) ⊆ a i P c (a i ) for i = 2, . .., N the plan is asymptotically stable.”

• AssumptionThe authors state this as an assumption: The model parameters, such as likelihood and transition matrices, are known.
• AssumptionThe authors state this as an assumption: The action templates with pre- and postconditions provided to the agent are correct.
• AssumptionThe authors state this as an assumption: A given desired goal is achievable by at least one atomic action.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The development effort of a prior node in a behavior tree is the same as that of a standard action node.

“Importantly, the development effort of a prior node in a BT is the same as a standard action node.”

May combine two claims.

Active inference uses a probabilistic state to filter spurious input: a bad observation merely lowers confidence slightly and does not cause a wrong transition.

“In active inference, the probabilistic representation of a state helps filter out this kind of spurious sensory input.For instance, a wrong observation like the one above would have caused the robot to be just slightly less confident about that state, with no erroneous transition.”

Classified as probably not a substantial claim. May combine two claims.

The lemma is a consequence of composing FTS actions sequentially: each action's effects fall inside the next action's attraction domain, and each controller's stable equilibrium is either C or inside an earlier action's region.

“Lemma 1 is a direct consequence of the sequential behavior composition of FTS actions, where each action has effects within the domain of attraction of the action below.The asymptotically stable equilibrium of each controller is either the goal C, or it is within the region of attraction of another action earlier in the sequence, see [3], [39], [40].”

• cites [3] sourceNot checked — no full text of the source
• cites [39] sourceNot checked — no full text of the source
• cites [40] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable).

Offline planning is better suited to offloading runtime computation for the parts of a task that do not change frequently.

“offline planning is arguably better suited to offload computations at runtime for the parts of the task that do not change frequently.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Replanning a small subtask at the same frequency as the active inference node can achieve reactiveness similar to the proposed approach.

“By replanning for a small subtask at the same frequency used for the active inference node, one can achieve similar reactiveness to our approach.”