Theatre-in-the-Loop: AI in Dramatic Practice
- Theatre-in-the-Loop is a family of methodologies that integrates theatrical practice into system design through dramaturgy, curation, and embodied performance.
- It employs mixed-initiative control with layered constraints, blending AI-generated outputs with human oversight to shape dynamic narratives.
- Applications span automated playwriting, live improv systems, expressive robotics, and educational tools, evaluated through public staging and performance metrics.
Taken together, the literature suggests that Theatre-in-the-Loop denotes a family of methodologies in which theatre is not treated as a mere downstream application area for AI, robotics, or HCI, but as the operative loop through which systems are planned, constrained, rehearsed, curated, embodied, and evaluated. In this framing, dramaturgical structure, performer expertise, directorial intent, stage conditions, and audience-facing performance become constitutive parts of the system design. The term is explicit in some works and implicit in others, but across automatic playwriting, live improvised dialogue systems, interactive drama, mixed reality staging, expressive robotics, educational chatbots, and accessibility pipelines, the recurrent pattern is a closed or semi-closed loop between machine generation and theatrical practice (Rosa et al., 2020, Mathewson et al., 2018, Panagiotidis et al., 5 Aug 2025).
1. Genealogy and conceptual range
Early computational theatre systems approached the loop from different entry points. Improbotics embedded a sequence-to-sequence recurrent neural network chatbot into live improvised theatre, with human controllers curating candidate lines and performers justifying them physically and emotionally in front of an audience (Mathewson et al., 2018). THEaiTRE and THEaiTRE 1.0 approached the problem from script generation, proposing hierarchical generation from title and synopsis to full play, with theatre experts defining prompts, reranking outputs, and finally realizing the production (Rosa et al., 2020, Rosa et al., 2021). Towards Intelligent Interactive Theatre described an intelligent agent that is simultaneously a drama manager and an onstage actor, with the audience treated as observer and optimization target rather than interactive player (Velissaris et al., 2019).
Later work broadened both the technical substrate and the meaning of the loop. Dramatron incorporated theatre and film professionals directly into prompt-chained script generation and staged resulting works at the Edmonton International Fringe (Mirowski et al., 2022). Designing and Evaluating Dialogue LLMs for Co-Creative Improvised Theatre turned a month-long Edinburgh Festival Fringe run into a live engineering cycle for multi-party LLMs (Branch et al., 2024). Theatre in the Loop: A Rehearsal-Based, Collaborative Workflow for Expressive Robotic Behaviours made theatre practice itself—directing, dramaturgy, and embodied improvisation—the generative core of expressive robot behavior design, explicitly distinguishing this from generic human-in-the-loop or artist-in-the-loop paradigms (Panagiotidis et al., 5 Aug 2025).
The concept also extends beyond conventional stage production. In A Piece of Theatre, teachers describe themselves as playwrights and directors when authoring LLM chatbot role-play for cyberbullying education (Hedderich et al., 2024). In TS-RGBD Dataset, the theatre environment becomes part of an assistive loop for visually impaired audiences, combining RGB, depth, and skeleton modalities for scene description and human action recognition (Benhamida et al., 2023). In Open-Theatre and From Role-Play to Drama-Interaction, the loop is formalized as interactive drama with scene graphs, triggers, memory, and multi-agent orchestration (Xu et al., 20 Sep 2025, Wu et al., 2024). This suggests that Theatre-in-the-Loop is best understood not as a single architecture but as a methodological family unified by the embedding of theatrical process into computation.
2. Constituent roles and loop structure
Across the literature, the loop is organized around recurring human and machine roles: playwrights or dramaturges define narrative constraints, operators or curators mediate runtime outputs, performers or puppeteers embody generated material, and audiences function either as evaluators, observers, or indirect supervision signals. The machine side may be a LLM, a multi-agent drama system, a motion-capture avatar stack, a robot playback pipeline, or a multimodal perception system. What distinguishes the paradigm is not the specific model class but the fact that theatre practice supplies the control logic.
| Instantiation | Human roles in the loop | Technical substrate |
|---|---|---|
| THEaiTRE / THEaiTRE 1.0 | theatre experts, dramaturge, director | GPT-2-style or GPT-2 XL script generation, summarization, MT (Rosa et al., 2020, Rosa et al., 2021) |
| Improbotics / Fringe improv systems | CEO, Puppet Master, Operator, Curator, Cyborg, improvisers | seq2seq RNN or multiple LLMs, ASR, TTS, headphones (Mathewson et al., 2018, Branch et al., 2024) |
| Interactive drama toolkits | player, developer, monitor, director-like controller | scene graphs, triggers, Narrative Chain, hierarchical memory (Wu et al., 2024, Xu et al., 20 Sep 2025) |
| Expressive robotics / mixed reality | director, puppeteer, roboticists, designers, manipulactor, mocaptor | Freedrive or gravity-compensation capture, joint-space replay, avatar control (Panagiotidis et al., 5 Aug 2025, Gagneré et al., 2023) |
| Educational and accessibility systems | teacher, student, operator, user | prompt-chained dialogue trees, RGB-D captioning, skeleton HAR (Hedderich et al., 2024, Benhamida et al., 2023) |
A common misconception is that Theatre-in-the-Loop is simply a theatrical instance of human-in-the-loop. The robotics formulation makes the distinction most sharply: theatre practice itself—directing, dramaturgy, embodied improvisation, rehearsal cycles—becomes the generative engine of behavior creation, rather than an external source of labels or corrections (Panagiotidis et al., 5 Aug 2025). Comparable logic appears in script systems where the dramaturge’s accept/reject decisions and scene assembly are part of the generation mechanism, not merely editorial post-processing (Rosa et al., 2021).
Another recurring structural feature is partial rather than total automation. In several systems, machine generation is deliberately narrowed to a tractable layer. THEaiTRE 1.0 generates line-by-line scene continuations that are later assembled by a dramaturge (Rosa et al., 2021). Improbotics constrains agency by routing all machine lines through a CEO Controller and preserving narrative coherence through Free-will Humans (Mathewson et al., 2018). Open-Theatre formalizes a comparable division through developer, player, and monitor consoles, plus director-style orchestration agents (Xu et al., 20 Sep 2025).
3. Planning, memory, and controllability
The script-generation branch of Theatre-in-the-Loop is dominated by hierarchical control. THEaiTRE proposes a pipeline of title or high-level prompt → textual synopsis → full play, with the play converted into a common JSON structure of acts, scenes, and dialogues. Character lists are generated explicitly and then enforced during dialogue generation; stage directions are mined from scenic remarks or settings and carried forward as scene-level guidance (Rosa et al., 2020). The formal core is standard autoregressive language modeling,
paired with conditional generation under an outline,
where contains title, synopsis, character list, and act or scene goals (Rosa et al., 2020).
THEaiTRE 1.0 operationalized a simpler but concrete version of this control logic with a vanilla GPT-2 XL used without fine-tuning. It enforced the set of allowed characters by modifying the next-token probability distribution at the start of each line, raised the repetition penalty from $1.00$ to $1.01$, and managed the $1024$-token context window through TextRank summarization with , , and (Rosa et al., 2021). Dramatron introduced a related hierarchy—log line → title → characters → plot outline → location descriptions → dialogue—and added previous-scene context to dialogue prompts after practitioner feedback, alongside nucleus sampling with , temperature 0, and loop detection (Mirowski et al., 2022).
Interactive-drama systems push controllability from hierarchical prompting toward explicit scene-state models. From Role-Play to Drama-Interaction represents drama as a directed graph of scenes with sub-narratives organized as a Narrative Chain, runtime state
1
and trigger-driven progression constrained by plot, character, and interaction conditions (Wu et al., 2024). Open-Theatre instead centers a hierarchical retrieval memory with Global, Event, Summary, and Archive stores. For candidate memory entry 2, retrieval is scored by
3
where recency is decayed differently across scenes and within the current scene (Xu et al., 20 Sep 2025). This architecture turns long-form dramatic coherence into an explicit retrieval and orchestration problem rather than relying entirely on prompt-local context.
Across these systems, controllability is achieved less by fully specified symbolic drama managers than by layered constraints: character gating, outline conditioning, scene-by-scene prompting, trigger matching, memory retrieval, and selective human reranking. A plausible implication is that Theatre-in-the-Loop has converged on mixed-initiative control: abstract narrative planning remains externally scaffolded, while local realization is delegated to generative models.
4. Embodied runtime systems and interfaces
Live improv systems expose the most stringent timing demands. In Improbotics, the CEO Controller types scene context into a web interface, the chatbot generates 10 candidate sentences word-by-word, the top 4 are ranked by LLM log-likelihood, and the CEO selects one or more suggestions for TTS delivery over FM to performers’ headphones (Mathewson et al., 2018). The system’s median response time exceeds 2 seconds, with some responses taking up to 4 seconds, so performers compensate through physicality, nonverbal action, and selective skipping of outdated lines (Mathewson et al., 2018). The loop therefore includes not just model inference and human curation, but embodied latency masking.
The Edinburgh Festival Fringe LLM deployment made this runtime stack more elaborate but structurally similar. Three models—ChatGPT-3.5, PaLM 2, and a locally served 4-bit quantized Llama 2 13B—ran concurrently; Whisper via whisper.cpp provided ASR with typical latency of ~0.3s; an Operator injected speaker names and scene metadata; and a Curator selected lines from a continuous AI stream for TTS delivery to a Cyborg performer (Branch et al., 2024). The system’s technical purpose was not only to generate text but to maintain multi-party timing, addressee resolution, and scene-relevant style through prompt buttons such as “more snarky” and “more punny” (Branch et al., 2024).
Robotics variants replace text-generation latency with capture and replay fidelity. In the rehearsal-based expressive robotics framework, a director provides layered inner monologue cues, a puppeteer manipulates a Franka Emika Panda in gravity-compensation or a UR3e in Freedrive, and roboticists log joint trajectories as time-stamped positions 4 for later replay via ROS 2’s FollowJointTrajectory (Panagiotidis et al., 5 Aug 2025). No smoothing, optimization, or learned dynamical primitives were applied in the reported trials; preserving live improvisation’s microdynamics was itself treated as an expressive design decision (Panagiotidis et al., 5 Aug 2025). Mixed-reality work in AvatarStaging shows an allied but distinct runtime loop, with mocaptors driving local avatar pose, manipulactors adjusting global transform, and directors managing scenic address and eyelines across physical and virtual spaces (Gagneré et al., 2023).
These embodiments clarify that Theatre-in-the-Loop is not limited to textual co-creation. It encompasses any pipeline in which theatrical rehearsal and performance constraints directly shape sensing, control, capture, or actuation. In that sense, the “loop” may be linguistic, dramaturgical, kinematic, or perceptual.
5. Evaluation through rehearsal, staging, and audience exposure
A defining feature of the paradigm is that evaluation is frequently inseparable from production. THEaiTRE identifies human expert assessment as central because automatic metrics are “not specified” and dramaturgical quality is difficult to measure automatically; theatre professionals pick or rerank outputs and ultimately stage the play, while audience reception at public premieres is treated as real-world evaluation of narrative quality and performability (Rosa et al., 2020). THEaiTRE 1.0 similarly reports practical success through production of a staged play, with the authors estimating that over 90% of the final staged script text originated from the automated tool (Rosa et al., 2021).
Dramatron made this coupling between system study and theatre practice especially explicit. It involved 15 paid theatre and film professionals in co-writing sessions and staged 5 co-written works in 7 performances at the Edmonton International Fringe (Mirowski et al., 2022). Survey results report 92% agreement on surprise and 77% agreement on enjoyment among respondents who completed the Likert survey, while creative-team reflections emphasize how repetition and “glitch” could become usable performance material rather than merely defects (Mirowski et al., 2022). The evaluation object is therefore not only script coherence but stage affordance.
Improbotics and the Edinburgh LLM study evaluate runtime co-creation more directly. Improbotics used post-show questionnaires adapted from the Presence Questionnaire and audience guessing in a Turing-test-like format; rehearsal improved performer proficiency and reduced perceived interface interference, with London performers reporting the highest proficiency and Edmonton performers the greatest difficulty focusing on performance rather than mechanisms (Mathewson et al., 2018). The Edinburgh deployment collected audience data from n=150, actor data from n=21, and in-situ observations across 26 consecutive shows and 1750+ attendees, showing sustained public interest alongside modest perceived naturalness and persistent recognition of machine limitations (Branch et al., 2024).
More infrastructure-oriented systems introduce proxy metrics. Open-Theatre evaluates memory and orchestration with Retrieval Accuracy, Response Plausibility, Narrative Coherence, Character Consistency, Multi-Agent Coordination, Plot Adherence, latency, and LLM calls per turn, reporting that memory improves quality across architectures and that Director–Global Actor offers a favorable trade-off between coherence and efficiency (Xu et al., 20 Sep 2025). Accessibility-oriented work on TS-RGBD evaluates dense captioning with BLEU, METEOR, ROUGE-L, and CIDEr, and reports 97.5% accuracy for egocentric direction on TS-RGBD images versus 89% on Visual Genome images with AdaBins depth estimation (Benhamida et al., 2023). This suggests that Theatre-in-the-Loop can also be evaluated through task metrics when theatre functions as an instrumented environment rather than a site of live dialogue generation.
6. Limitations, misconceptions, and future trajectories
The literature repeatedly rejects the idea that Theatre-in-the-Loop already delivers autonomous theatrical intelligence. Script systems report missing plot structure, lack of progression toward scene conclusions, random character introductions, degenerate repetition, and weak long-range character motivation (Rosa et al., 2020). Live improv systems report latency, ASR errors, limited fine-grained context tracking, persona inconsistency, and the need for continual human moderation to prevent offensive or irrelevant outputs (Mathewson et al., 2018, Branch et al., 2024). Robotics variants expose mechanical interruptions such as abrupt joint locking and halts on the Franka platform, as well as the broader problem of aligning expressive intent with platform affordances (Panagiotidis et al., 5 Aug 2025).
Another misconception is that audience interactivity is necessary. Some systems make the audience an explicit participant or evaluator, but others optimize for audience experience while keeping the audience observational. The intelligent interactive theatre formulation is explicit on this point: the audience is an observer, and the drama manager seeks a satisfying, unpredictable narrative under authorial constraints (Velissaris et al., 2019). Conversely, educational and interactive-drama systems shift participation toward students or players, making the loop closer to role-play than to spectatorship (Hedderich et al., 2024, Wu et al., 2024).
Safety, bias, and cultural context remain weakly resolved. Improbotics notes biases inherited from movie-subtitle training data and the need for human moderation (Mathewson et al., 2018). THEaiTRE states that cultural context, safety, and bias are not explicitly addressed, though domain-sensitive fine-tuning and human oversight may mitigate them (Rosa et al., 2020). Educational chatbot work highlights the need for guardrails so the system does not endorse harmful behaviors, and asks for higher-level controls over repetition, topic pivots, and refusal behavior (Hedderich et al., 2024).
The main research trajectory is toward richer controllability plus embodiment, rather than toward removal of the human. Proposed directions include larger theatre-specialized models, stronger entity tracking and adapted coreference, richer scene and beat representations, automated moderation, multimodal grounding, better turn-taking policies, rehearsal-derived online learning, dynamic reconfiguration of robot gestures, and open-source infrastructures for reproducible interactive drama (Rosa et al., 2020, Branch et al., 2024, Panagiotidis et al., 5 Aug 2025, Xu et al., 20 Sep 2025). The generalized blueprint is already visible: collect theatre-grounded data, represent scripts or scenes in structured form, impose hierarchical or retrieval-based control, preserve a human curation layer, and evaluate through rehearsal, staging, and audience-facing deployment. Under that blueprint, Theatre-in-the-Loop is less a finished technique than a durable research program for making theatrical process itself part of computational intelligence.