This research aims to bridge the gap between symbolic plans and executable, constraintbased robot task specifications by means of offline skill programming, and online skill scheduling, refinement and adaptation. Symbolic plans are discrete sequences of actions that must be performed by a robot to achieve an intended change in the environment. Such actions are not directly executable, so a refinement process adds context-dependent knowledge (about the robot, its tools, and the environment) required to generate the robot’s continuous motion. This work enhances this refinement process, providing a declarative description of those situations that lie in both discrete and continuous domain, and a late online decision making about the symbolic and continuous representations of an action. For example, a multi-arm robot may have symbolic plans to open a drawer by means of the right arm, the left arm, or both, and it can delay its decision about which plan to execute until it knows all properties of the drawer and its environment. In the same vein, grasping the drawer’s handle requires the robotic gripper to be opened, and while that action can start anytime during the approach motion, the progress of both actions must be coordinated online, to avoid the robot to hit the handle with a not yet sufficiently opened gripper. The state-of-the-art solution is typically to first open the gripper, and only then start the approach motion; this work optimises the plan execution at runtime, by an execution engine that exploits the most current status of the robot capabilities. This behaviour is supported by formal models and Domain Specific Languages, and by mechanisms to specify, compose and coordinate robotic skills at runtime. A first contribution is a formal skill model that represents both discrete and continuous aspects of an action. The scheduling of the execution of these skills is constrained by a set of dependencies that must be satisfied online, and that represent logical conditions on the continuous state of the robot-environment interaction; the formal representation of these dependencies is called the Skill Dependency Graph (SDG). Such SDG models are interpreted and realised at runtime by an execution engine called Skill Dependency Graph Executive. This dissertation prefers declarative and formal representations instead of 1 procedural and code-based approaches: to this end, the micro Skill Dependency Language (USDL) is proposed as a minimal set of declarative rules behind SDG models. The second contribution of this dissertation is a novel constraint-based task specification based on formal, declarative geometric relationships to represent motion constraints; that is, relationships (e.g., distances and angles) between pairs of geometric entities (such as points, lines and planes). At runtime, these formal expressions are evaluated in the actual robot-environment context to generate explicit and instantaneous motion set-points, optimising the robot limitations (e.g., joint limits) and redundancies. Moreover, this solution decouples the motion specification from the underlying numerical solver; in this context, constraint-based optimal control strategies are preferred to constraint-based motion planning to provide reactiveness and to support specifications that describe physical interactions. The third contribution is to integrate the previous two, with the explicit aim to support generalization over skill models. Geometric entities are modelled as part of a geometric item description, whether it is a physical object, a virtual object or a tool (i.e., a controlled object that adds further capabilities to the robot). This modelling effort generalises the skill specification to a class of cases, where the objects involved are described with semantically equivalent and more generic geometric entities. For example, the skill that performs the grasping of a drawer’s handle conforms to the same generic skill as the grasping of a glass, since both handle and glass are approximated as a cylindrical envelope. This generalisation leads to skill prototypes, which are not executable as is, but require additional context-dependent knowledge: the output of this refinement is a executable skill instance. Complex skills are created by composition of multiple skill prototypes, preferred to inherit from a single skill; several experimental examples are provided in this work. Another contribution is the realisation of a Just-in-Time strategy, to postpone decisions until they are really needed, so that the task specifications are composed with the latest information available at runtime. A final contribution proposes hierarchical hypergraphs and a meta model called “NPC4” as formal representation of the structural aspects of graph-based models (including the SDG). Such a strict separation between structure and behaviour enables the reuse of infrastructure code, and simplifies the definition of new Domain Specific Languages.

Lo scopo di questa tesi di dottorato consiste nel trasformare piani di tipo simbolico in specifiche di movimento basate su vincoli ed eseguibili da sistemi robotici autonomi, attraverso la programmazione di skill di movimento, nonch`e la loro schedulazione e adattamento a tempo di esecuzione. I piani di tipo simbolico sono sequenze ordinate di azioni discrete eseguite da un robot atte ad ottenere un desiderato cambiamento sull’ambiente circostante. Le azioni, in quanto simboliche, non sono direttamente eseguibili, poich esse non contengono dettagli informativi che dipendono strettamente dal contesto di esecuzione dell’azione stessa (e.g., informazioni sulle caratteristiche del robot, degli strumenti del quale `e equipaggiato, nonch`e informazioni sull’ambiente circostante); per rendere eseguibili le azioni `e quindi necessario un processo di conversione di ciascuna azione simbolica in una skill che definisca come eseguire il movimento continuo del robot. Questa tesi propone una metodologia per formalizzare la suddetta trasformazione, fornendo una descrizione dichiarativa di quelle situazioni che esibiscono una natura sia discreta che continua e ritardando tale processo laddove possibile. Per esempio, un robot umanoide pu`o aprire un cassetto per mezzo del braccio destro, sinistro o entrambi: la scelta della strategia di apertura dev’essere ritardata fintanto che le informazioni siano insufficienti per una decisione adeguata. Inoltre, afferrare la maniglia di un cassetto richiede che la pinza robotica sia aperta: quest’ultima azione pu essere eseguita in qualsiasi momento durante il movimento di avvicinamento alla maniglia, ma evitando di urtare quest’ultima nel caso in cui la pinza non sia sufficientemente aperta. Una soluzione semplicistica `e quella di eseguire tali azioni in ordine strettamente sequenziale: questa tesi propone un’alternativa che ottimizza l’esecuzione del piano online, attraverso un motore d’esecuzione che sfrutta le caratteristiche del robot e le adatta al contesto di esecuzione. Un primo contributo della tesi `e la definizione di un modello formale di skill robotica che permetta di rappresentare un’azione in modo completo, considerandone gli aspetti nel dominio sia discreto che continuo. La schedulazione dell’esecuzione di queste skill `e vincolata ad un insieme di dipendenze da soddisfare a tempo di esecuzione; tali dipendenze rappresentano condizioni logiche relative alle interazioni tra robot e ambiente, le quali appartengono al dominio continuo. La rappresentazione formale di queste dipendenze 1 definisce un modello chiamato Skill Dependency Graph (SDG), che viene interpretato e realizzato a tempo di esecuzione da un Skill Dependency Graph Executive (SDG-E). Preferendo una descrizione dichiarativa e formale rispetto ad approcci procedurali, questa tesi propone inoltre il linguaggio micro Skill Dependency Language (USDL), come un insieme minimale di regole dichiarative che definiscono le relazioni in un modello SDG. Come secondo contributo, la tesi descrive un’innovativa specifica di movimento per mezzo di relazioni geometriche che ne rappresentano i vincoli, ovvero relazioni tra coppie di entit geometriche (punti, linee e piani) come distanze ed angoli. A tempo di esecuzione, queste espressioni formali sono valutate nel contesto attuale allo scopo di generare setpoints istantanei di movimento, ottimizzati rispetto ai limiti del robot (e.g., limiti di giunto) e ridondanza cinematica. Inoltre, questa soluzione separa la specifica di movimento dal risolutore numerico adottato; in tal senso, strategie di controllo ottimo su vincoli sono preferite alla pianificazione del movimento su vincoli, in quanto le prime supportano specifiche che includono interazioni fisiche o che implementano azioni di controllo reattive. Un terzo contributo `e derivato dall’integrazione dei due precedenti, avente lo scopo di generalizzare quelle componenti che formano un modello di skill programmato manualmente. Le entit geometriche discusse in precedenza sono modellate come parte di generici elementi geometrici, che siano oggetti fisici, oggetti virtuali o strumenti del robot stesso. Questo approccio permette di classificare le varie specifiche di skill robotiche, laddove gli oggetti coinvolti nella specifica stessa siano semanticamente equivalenti. Per esempio, la skill che realizza il movimento per afferrare una maniglia di un cassetto `e conforme alla stessa skill generica che afferra la maniglia di una porta o un bicchiere, in quanto questi oggetti possono essere approssimati come inviluppo di un cilindro. Una skill cos´ı generalizzata `e detta prototipo: skill prototipali non sono direttamente eseguibili, ma sono arricchite da un processo di trasformazione che provvede informazioni dipendenti dal contesto di esecuzione. Il risultato di tale processo converte il prototipo in un’istanza eseguibile in una skill di movimento. Skill complesse sono create per composizione di molteplici prototipi, piuttosto che applicando una strategia ad ereditariet`a da una sola skill. Questo lavoro di tesi provvede diversi esempi di skill implementate e testate sperimentalmente. Questa tesi contribuisce alla realizzazione di una strategia Just-in-Time, con lo scopo di ritardare ogni decisione fintanto che quest’ultime siano effettivamente necessarie per la realizzazione del movimento. Grazie a questa soluzione, la specifica di movimento `e composta a tempo di esecuzione considerando le ultime informazioni disponibili. Infine, questa tesi propone ipergrafi gerarchici e un meta modello chiamato “NPC4” per rappresentare formalmente aspetti strutturali di modelli basati su grafi (incluso il modello SDG). La separazione tra struttura e funzionalit`a permette un miglior riutilizzo di quelle porzioni di codice in comune, inoltre ne semplifica la definizione di nuovi linguaggi specifici di dominio.

Online Coordination and Composition of Robotic Skills: Formal Models for Context-aware Task Scheduling

SCIONI, Enea
2016

Abstract

This research aims to bridge the gap between symbolic plans and executable, constraintbased robot task specifications by means of offline skill programming, and online skill scheduling, refinement and adaptation. Symbolic plans are discrete sequences of actions that must be performed by a robot to achieve an intended change in the environment. Such actions are not directly executable, so a refinement process adds context-dependent knowledge (about the robot, its tools, and the environment) required to generate the robot’s continuous motion. This work enhances this refinement process, providing a declarative description of those situations that lie in both discrete and continuous domain, and a late online decision making about the symbolic and continuous representations of an action. For example, a multi-arm robot may have symbolic plans to open a drawer by means of the right arm, the left arm, or both, and it can delay its decision about which plan to execute until it knows all properties of the drawer and its environment. In the same vein, grasping the drawer’s handle requires the robotic gripper to be opened, and while that action can start anytime during the approach motion, the progress of both actions must be coordinated online, to avoid the robot to hit the handle with a not yet sufficiently opened gripper. The state-of-the-art solution is typically to first open the gripper, and only then start the approach motion; this work optimises the plan execution at runtime, by an execution engine that exploits the most current status of the robot capabilities. This behaviour is supported by formal models and Domain Specific Languages, and by mechanisms to specify, compose and coordinate robotic skills at runtime. A first contribution is a formal skill model that represents both discrete and continuous aspects of an action. The scheduling of the execution of these skills is constrained by a set of dependencies that must be satisfied online, and that represent logical conditions on the continuous state of the robot-environment interaction; the formal representation of these dependencies is called the Skill Dependency Graph (SDG). Such SDG models are interpreted and realised at runtime by an execution engine called Skill Dependency Graph Executive. This dissertation prefers declarative and formal representations instead of 1 procedural and code-based approaches: to this end, the micro Skill Dependency Language (USDL) is proposed as a minimal set of declarative rules behind SDG models. The second contribution of this dissertation is a novel constraint-based task specification based on formal, declarative geometric relationships to represent motion constraints; that is, relationships (e.g., distances and angles) between pairs of geometric entities (such as points, lines and planes). At runtime, these formal expressions are evaluated in the actual robot-environment context to generate explicit and instantaneous motion set-points, optimising the robot limitations (e.g., joint limits) and redundancies. Moreover, this solution decouples the motion specification from the underlying numerical solver; in this context, constraint-based optimal control strategies are preferred to constraint-based motion planning to provide reactiveness and to support specifications that describe physical interactions. The third contribution is to integrate the previous two, with the explicit aim to support generalization over skill models. Geometric entities are modelled as part of a geometric item description, whether it is a physical object, a virtual object or a tool (i.e., a controlled object that adds further capabilities to the robot). This modelling effort generalises the skill specification to a class of cases, where the objects involved are described with semantically equivalent and more generic geometric entities. For example, the skill that performs the grasping of a drawer’s handle conforms to the same generic skill as the grasping of a glass, since both handle and glass are approximated as a cylindrical envelope. This generalisation leads to skill prototypes, which are not executable as is, but require additional context-dependent knowledge: the output of this refinement is a executable skill instance. Complex skills are created by composition of multiple skill prototypes, preferred to inherit from a single skill; several experimental examples are provided in this work. Another contribution is the realisation of a Just-in-Time strategy, to postpone decisions until they are really needed, so that the task specifications are composed with the latest information available at runtime. A final contribution proposes hierarchical hypergraphs and a meta model called “NPC4” as formal representation of the structural aspects of graph-based models (including the SDG). Such a strict separation between structure and behaviour enables the reuse of infrastructure code, and simplifies the definition of new Domain Specific Languages.
BONFE', Marcello
TRILLO, Stefano
File in questo prodotto:
File Dimensione Formato  
tesi.pdf

accesso aperto

Tipologia: Tesi di dottorato
Licenza: PUBBLICO - Pubblico senza Copyright
Dimensione 33.45 MB
Formato Adobe PDF
33.45 MB Adobe PDF Visualizza/Apri

I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11392/2403213
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact