Embodied Playground Definition

• Embodied Playgrounds are environments where play, learning, or decision-making are organized through bodily interaction with spatially situated artifacts, computational systems, and social contexts.
• Key properties include embodied spaces that encourage social participation, affordances for physical tasks, coordinated activity through movements, and feedback-driven adaptations.
• Embodied Playgrounds have applications in digital twins for surgical training, multi-agent benchmarks for strategic decision-making, and mixed-reality sports for inclusive representations.

Embodied Playground is a design paradigm and research environment in which play, learning, or decision-making is organized through the interaction of bodies, spatially situated artifacts, computational systems, other agents, and social or environmental context. The term encompasses public tangible-interface arcades, multi-agent benchmarks, developmental robots, mixed-reality spaces, full-body virtual environments, digital twins, robot-learning simulators, inclusive sports representations, and long-horizon embodied-agent benchmarks. Across these settings, the central concern is not merely whether an agent or player receives visual information, but how bodily capabilities, action possibilities, physical consequences, spatial relations, social coordination, and feedback jointly constitute the activity.

1. Conceptual foundations and defining properties

An Embodied Playground treats the environment as more than a container for tasks. Its physical or simulated configuration, objects, interfaces, rules, spectators, agents, and feedback mechanisms mutually shape participation. The concept therefore differs from a conventional screen-centered game, static dataset, or isolated simulator. Its defining feature is the organization of action around situated embodiment.

Several properties recur across the research:

• Situated action: actions acquire meaning from location, orientation, objects, obstacles, resources, and the state of other agents.
• Material or virtual affordances: objects are represented by what they permit an agent to do, not only by appearance or semantic category. Affordances may include grasping, pushing, pulling, lifting, turning, containing, or supporting; situatedness captures the transient relation between object and agent (Xu et al., 2024).
• Bodily participation: movement, posture, exertion, proprioception, manipulation, or body orientation functions as an input, output, or both.
• Intercorporeality: participants coordinate through one another’s movements, attention, proximity, timing, and responses.
• Environmental agency: walls, floors, furniture, terrain, lighting, sound, virtual fields, or dynamic equipment become active constraints or resources.
• Feedback-driven adaptation: actions alter the environment, whose consequences inform subsequent decisions.
• Open-ended experimentation: many Embodied Playgrounds emphasize improvisation, unusual controls, multiple solutions, or reversible configuration rather than a single optimal action sequence.
• Social visibility: spectators, bystanders, teammates, coaches, facilitators, or simulated social characters may influence the activity and its interpretation.

The paradigm includes both human-centered and machine-centered forms. In Rabble Room Arcade, unusual physical interfaces encouraged active, co-located, and public play (Gavin et al., 2016). In Pommerman, spatially constrained actions, delayed hazards, resource management, partial observability, cooperation, and competition form a multi-agent strategic playground (Resnick et al., 2018). In robot-learning systems such as MuJoCo Playground, embodiment includes morphology, sensing, actuation, physics, rewards, and the hardware interface connecting simulation to deployment (Zakka et al., 12 Feb 2025).

The term is consequently broader than “embodied interaction.” It denotes not only a mode of interaction but an environment, infrastructure, or ecology in which embodied interaction can be explored, learned, evaluated, or socially negotiated.

2. Bodily, material, and perceptual embodiment

Tangible and physically expressive interaction

Embodied Playgrounds often reject the assumption that interfaces should minimize effort, maximize speed, or disappear from awareness. Rabble Room Arcade deliberately used inefficient interfaces—including shovels, cranks, wooden sliders, tilt boards, foot-operated buttons, a toilet-chain switch, rotary encoders, and body silhouettes—to produce spectacle, humor, coordination, and participation (Gavin et al., 2016). Its eight games included cooperative, competitive, and experimental formats such as Shadow Showdown, Elevator, Double Shovel, Word Wars, and Fruit Racers. The event attracted more than 100 visitors, while six of the eight games involved external contributors.

The resulting interaction was embodied because players pushed, pulled, tilted, cranked, shoveled, stepped, posed, and coordinated. It was playground-like because awkwardness and novelty were treated as meaningful features rather than defects. The approach also challenged the assumption that embodiment requires sophisticated sensing or large bodily movement: Word Wars used one button per player, yet its shared cabinet and multiplayer structure produced a “tight social experience.”

The body as play material

Auto-Paízo Games extends embodiment beyond the body-as-controller model. In the three games, one hand provided voluntary input while Electrical Muscle Stimulation actuated the other hand as a computer-controlled opponent (Patibanda et al., 2023). The system thus made the body both input and output. Participants described “loaning” control to the computer, while the EMS hand could be experienced as an opponent, companion, robot, separate person, or part of the self.

The seven-day field study involved 12 participants, although only 11 successfully played the games. Participants appreciated varied bodily movement and ambiguity, but EMS calibration, electrode replacement, discomfort, and actuation speed remained significant constraints. Slap-Me-If-You-Can, which required gross-motor movement, was particularly difficult for five participants because the EMS hand or arm did not move quickly enough. The findings support a design tension central to Embodied Playgrounds: bodily uncertainty can encourage prediction and exploration, but excessive uncertainty can undermine fairness, agency, comfort, and safety.

Mixed-reality bodily movement

Ball-AR used augmented reality to turn a shared room into a dodgeball arena (Kumar et al., 2023). Two players scanned the room, threw virtual balls, used physical surfaces for bounces, ran to retrieve balls, and dodged one another. Its mechanics linked movement to meaningful game consequences rather than treating physical activity as an ornamental addition. Participants reported memorable experiences, direct social reactions, and physical urgency. However, detected surfaces functioned primarily as strategic resources; they did not consistently produce a strong sense of shared place. Participants also wanted more virtual objects and richer augmentation.

In public-space mixed reality, MOFA—Multiplayer Omnipresent Fighting Arena—used see-through head-mounted displays, Apple Watch gesture sensing, physical wands, peer-to-peer networking, and handheld spectator views to support The Training, The Duel, and The Dragon (Hu et al., 18 May 2025). The three probes enabled single-player practice, competitive two-player combat, and asymmetric play involving headset-wearing warriors and a handheld dragon puppeteer. The study involved 65 players across 20 locations and found that bodily movement, spectatorship, physical props, and multiplayer participation strongly shaped social acceptance. At the same time, safety, privacy, awkwardness, exclusion of non-users, and the mismatch between players’ immersive views and bystanders’ perspectives remained unresolved.

GravField develops a less prescriptive model of mixed-reality movement (Hu et al., 17 Sep 2025). Its Rope, Spring, and Magnetic Field objects used configurable digital physics, audiovisual feedback, and live facilitation to nudge collective movement without prescribing choreography. Participants developed synchronization, mirroring, relay patterns, orbital movement, push–pull interactions, and improvised interpretations. The Object Jockey could change mass, elasticity, friction, force, field strength, and mappings between movement and sound. The system demonstrates how an Embodied Playground can guide without directly controlling, provided that reconfiguration remains transparent enough to preserve agency.

Full-body narrative and immersive environments

The Magical Experience Generator combines a Torus Treadmill, an immersive cuboid projection environment, body and hand tracking, and a Unity forest based on “The Man Who Made Flowers Bloom” (Ye et al., 2024). Participants physically walk, grab virtual ashes, throw them using a tracked controller, and observe withered trees gradually bloom. The installation uses a $4 \times 4 \times 2.25$ m display space, eight projectors, and a treadmill with a maximum walking speed of $0.6$ m/s. The system demonstrates a narrative form of embodiment in which a participant physically enacts the story’s central transformation. Its evaluation, however, is an implementation test based on trajectory recording rather than a controlled user study.

Crossmodal language grounding provides a developmental interpretation of the playground (Heinrich et al., 2020). In the EMIL scenario, the humanoid robot NICO manipulates toy-like objects through pushing, pulling, lifting, and scooting while receiving caregiver descriptions. Auditory, visual, and sensorimotor sequences are integrated by adaptive multiple-timescale recurrent neural networks. The model suggests a pathway from embodied interaction to language: temporally extended bodily episodes become crossmodal abstractions that are decomposed into words or phonemes. The reported results show that adaptive temporal hierarchy improves complete-description accuracy, although the dataset remains small, supervised, and highly constrained.

3. Strategic and social multi-agent environments

Embodied Playgrounds need not involve human bodies directly. They may instantiate strategic embodiment through spatially situated agents whose actions are constrained by objects, time, resources, hazards, and other agents.

Pommerman is a canonical example. Each episode begins on a randomly drawn symmetric $11\times11$ grid with four agents, rigid and destructible walls, passages, bombs, and power-ups (Resnick et al., 2018). Agents have six discrete actions: Stop, Up, Left, Down, Right, and Bomb. Bombs explode after ten time steps, destroy wooden walls and objects within blast reach, and can trigger chains. Some agents can kick bombs, turning them into mobile projectiles.

The environment includes free-for-all and two-versus-two team scenarios, with optional communication and full or partial observability. In partially observable variants, agents see a $5\times5$ purview while other board cells are replaced by fog. Communicative agents send two words from an eight-word dictionary, with a one-step delay. The resulting tasks require planning, opponent and teammate modeling, resource management, game-theoretic reasoning, communication, adaptation to previously unseen teammates, and coordination under delayed consequences.

Pommerman’s significance lies in its strategic embodiment. It omits continuous control, realistic body morphology, three-dimensional physics, raw visual perception, and sensor noise, but preserves action constraints arising from position, objects, time, resources, hazards, and social interaction. It therefore represents a narrow but technically tractable form of embodied decision-making.

The same principle appears in EmboMatrix, a scalable training ground for embodied decision-making (Lei et al., 14 Oct 2025). EmboMatrix generates tasks and scenes, executes LLM-produced symbolic robot skills, computes structured rewards, and updates the model through interactive rollouts. Its 45 base scenes are expanded through multi-agent social simulation, multi-level layout generation, heterogeneous simulation workers, and a hierarchical reward architecture. The platform supports 13 action primitives, including moving, turning, picking up, placing, opening, closing, toggling, heating, cooking, freezing, and moving to a room.

EmboMatrix’s rewards progress from valid syntax to semantic relevance to goal satisfaction. EmboBrain-7B achieved 65.8% on Internal-Verified and 62.9% on the EAI benchmark, surpassing the 671B DeepSeek-R1 baseline by an average of approximately 9.5 percentage points across the two reported benchmarks. These results support the importance of environment-grounded interaction, although the experiments remain simulated, use a narrow skill vocabulary, and do not establish physical-robot transfer.

The newly introduced Embodied Playground benchmark formalizes the long-horizon version of this problem (Chen et al., 19 Oct 2026). It contains 410 task environments spanning launcher scenes, Tower of Hanoi, pocket cube, Rubik’s cube, Kongming lock, bookshelf tasks, and ManiSkill household activities. EP-100 selects 100 tasks across five suites and 16 task types; EP-Easy and EP-Hard provide fixed 50-task splits; EP-20 is a compact 20-task diagnostic selection. The benchmark evaluates whether foundation-model agents can track task progress, satisfy intermediate constraints, revise plans through execution feedback, respect resource budgets, and maintain physical validity across long horizons.

Its task families combine strategic, spatial, and physical demands. Launcher tasks involve aiming, indirect causal effects, fracture, finite ammunition, and protected targets. Tower tasks require legal rearrangement, grasping, alignment, and stable placement. Kongming-lock tasks require ordered, contact-rich assembly and disassembly. Cube tasks require state tracking and prescribed physical turns. Home tasks include sorting, insertion, hooking, interception, pouring, and multi-stage household activities. This benchmark treats the Embodied Playground as a unified evaluation space for reasoning, tool use, physical execution, and state management.

4. Computational training grounds and digital twins

GPU-native robot-learning environments

MuJoCo Playground integrates MJX physics, the Madrona batch renderer, training environments, robot assets, domain randomization, and deployment interfaces (Zakka et al., 12 Feb 2025). It supports quadrupeds, humanoids, dexterous hands, robot arms, state observations, and pixel observations. The framework simulates thousands of environments in parallel and supports PPO, SAC, Brax, and RSL-RL workflows.

Its embodied scope includes morphology, joint and actuator behavior, sensing, contact geometry, low-level control, rewards, termination conditions, and hardware interfaces. The framework reports configurations of 2,048, 8,192, and up to 32,768 parallel environments. It demonstrates zero-shot transfer for quadruped locomotion, humanoid walking, dexterous cube reorientation, non-prehensile manipulation, and RGB-based cube grasping. Its limitations include compilation overhead, contact-scaling constraints, visual-domain dependence, and the need for accurate robot, camera, actuator, and timing models.

MO-Playground extends this infrastructure to multi-objective reinforcement learning (Janwani et al., 10 Mar 2026). It uses MORLAX, a GPU-compatible actor–critic method with hypernetworks that generate preference-conditioned policies. The environment returns vector-valued rewards rather than a single scalar, allowing a policy family to trade off objectives such as tracking, energy efficiency, smoothness, arm swinging, and rigidity.

The BRUCE humanoid experiment optimized six objectives and produced a continuous family of policies. One learned policy emphasized pronounced arm swinging, another nearly stationary arms, and a third maximum smoothness. MORLAX achieved 21–270 times the speed of HYPER-MORL across benchmark environments while attaining higher reported hypervolumes. The framework’s principal limitation is that linear scalarization primarily recovers supported convex portions of the Pareto front, while objectives such as comfort, naturalness, or social acceptability remain difficult to encode.

Vision-centric simulation and Real2Sim

GS-Playground combines a parallel physics engine, batched 3D Gaussian Splatting renderer, and automated Real2Sim pipeline (Jia et al., 28 Apr 2026). Its purpose is to make photorealistic vision-informed reinforcement learning feasible at large scale. The system binds Gaussian primitives to rigid bodies through Rigid-Link Gaussian Kinematics, allowing physics to remain the source of motion and 3DGS to provide visual appearance. It supports RGB, depth, LiDAR, height scans, force and torque sensing, and contact information.

The reported rendering throughput is approximately $10^4$ FPS at $640\times480$, with up to 2,048 simultaneous scenes. The system also reconstructs scenes and objects from RGB images using Grounding DINO, SAM1/SAM2, AnySplat, SAM-3D, inpainting, alignment, pruning, and collision-geometry integration. A key manipulation experiment achieved 90% zero-shot success in 20 real-world trials for RGB-based cube picking, compared with 0% for the reported MuJoCo, ManiSkill3, and Isaac Lab comparison policies under that experiment.

TwinOR applies the digital-twin paradigm to dynamic operating rooms (Zhang et al., 10 Nov 2025). It reconstructs static geometry from pre-scan video and dynamic human and equipment motion from synchronized multi-view observations. The system uses Neuralangelo for room surfaces, SMPL for human bodies, ViTPose+, SAM2, stereo point clouds, RANSAC, FPFH, colored ICP, and Blender fusion. It achieves centimeter-level room reconstruction, with reported mean Chamfer distances of 14.1 mm for OR-A and 22.7 mm for OR-B.

TwinOR supports photorealistic rendering, viewpoint synthesis, workflow replay, sensor simulation, spatial navigation, human–robot coordination, and perception benchmarking. FoundationStereo and ORB-SLAM3 operated on TwinOR-generated data with reported results within the range of real indoor benchmarks. However, TwinOR is primarily a visual and kinematic twin rather than a validated physics-rich surgical simulator. Fine contact, tissue deformation, deformable instruments, real-time updates, and comprehensive sim-to-real validation remain open problems.

Alexa Arena provides a user-centric simulation platform for interactive embodied AI (Gao et al., 2023). It contains 10 multi-room layouts, 336 unique objects, 14 object affordance properties, configurable missions through the Challenge Definition Format, real-time web interaction, dialogue, game-like guidance, and human-generated instructions and question–answer pairs. Its benchmark includes approximately 3,000 missions, 46,563 instructions, and 46,233 question–answer pairs.

The strongest reported baseline achieved 34.20% mission success with question–answer information. A user study with 31 non-expert participants reported an average mission success rate of 81.72% and a mean ease-of-learning score of 6.10 on a seven-point scale. The platform illustrates a human-centered Embodied Playground in which interaction data, instruction following, dialogue, visual grounding, navigation, manipulation, and causal state changes are studied together.

5. Representation, learning, and inclusivity

Objects as agent–object relations

The proposal that object space is embodied has direct implications for Embodied Playgrounds (Xu et al., 2024). Familiar mugs and novel objects were judged not only by visual properties but also by action atoms describing affordances and situatedness. In the familiar-object experiment, combined embodied features explained an adjusted $R^2$ of 0.57 in subjective dissimilarity, while visual features alone explained an adjusted $R^2$ of 0.17. Novel objects produced a corresponding adjusted $R^2$ of 0.42 for embodied features.

These results support an agent-specific representation in which an object is characterized by what a particular body can do with it in a particular context. A mug’s representation changes with hand orientation, distance, object orientation, grip, and movement direction. An Embodied Playground should therefore model objects not merely as meshes or class labels, but as dynamic sets of possible interactions conditioned on body, goal, and situation.

Inclusive transformation of embodied knowledge

BRIDGE applies embodiment transformation to wheelchair basketball (Saiki et al., 26 Feb 2026). It reconstructs broadcast basketball footage and maps standing-player movement into wheelchair-specific representations by separately modeling wheelchair base, trunk, and head orientation. The base captures mobility direction, the trunk encodes posture and functional class, and the head represents attention and tactical intent.

Across 80 comparisons, mapped representations were preferred over a movement-only baseline 66 times, and nearly 90% of ratings judged mapped postures natural. In a second study, wheelchair-reconstructed footage significantly improved self-efficacy, imagery vividness, ease of imagery, perceived tactical usefulness, exploration of new tactics, and intention to use. The work demonstrates that accessibility can involve transforming the bodily assumptions of content itself, rather than merely adding an interface accommodation.

BRIDGE also exposes the limits of representational transformation. Its evaluation is primarily observational, and it does not demonstrate improved on-court performance, long-term retention, direct avatar control, collision simulation, or physical practice. The system should therefore be understood as an inclusive visual learning and coaching environment rather than a complete sports simulator.

Emotional and therapeutic embodiment

“Tangible Intangibles” proposes an embodied playground for emotional expression using clay, drawing, biometrics, and mixed reality (Nasri et al., 15 Dec 2025). Participants begin with somatic grounding and manual art-making, then explore MR-based biometric journaling in which breath, heart-rate variability, eye-movement dynamics, and body movement influence color, pulsing, motion quality, and spatial form.

The proposal treats biosignals as expressive material rather than objective emotional truth. Its intended loop is:

$$\text{body} \rightarrow \text{sensors} \rightarrow \text{visualization} \rightarrow \text{awareness} \rightarrow \text{body}.$$

The work is a workshop and prototype proposal rather than a completed empirical study. It reports no participant sample, signal-processing specification, classification accuracy, statistical analysis, or validated therapeutic outcome. Its significance lies in the design tension between automated translation and personal interpretation. A trauma-informed Embodied Playground would require intensity controls, opt-out mechanisms, privacy protection, alternative modalities, and clear separation between expressive visualization and diagnosis.

Theatrical and compositional environments

Devising Interactive Spaces applies the playground model to immersive theatre (Panagiotidis et al., 7 Jul 2026). Across six workshops with eight professional performance-makers, gesture, position, and speech recognition were connected through Node-RED to Philips Hue lighting and sound playback. Participants created light-and-sound scores, responsive architectures, participatory prototypes, and a multi-room scratch performance.

The system’s key contribution is not sensing accuracy but compositional availability. Visible mappings, quick configuration, modular inputs and outputs, and immediate in-room testing allowed performers to treat sensing and actuation as rehearsal materials. The system is therefore better understood as a creative prototyping tool than a production-grade show-control platform. Its limitations include approximate calibration, fragile recognition, limited speech fidelity, modest output control, and continued dependence on technical support.

6. Evaluation, limitations, and future directions

Evaluation across Embodied Playgrounds varies substantially by purpose. Public and artistic systems commonly use observational and qualitative methods. Rabble Room Arcade relied primarily on descriptive observations, attendance, external participation, and organizer reflection (Gavin et al., 2016). Ball-AR used five dyads and semi-structured interviews without quantitative movement logging or inferential tests (Kumar et al., 2023). Auto-Paízo used thematic analysis of 704 interview data units, with inter-rater reliability of 97.43% and Cohen’s $0.6$0 (Patibanda et al., 2023). MOFA combined in-the-wild observations, interviews, spectator videos, field notes, and online comments, but its participant sample was opportunistic and its conclusions exploratory (Hu et al., 18 May 2025).

Simulation and benchmark environments support more formal metrics. Pommerman uses win rates, survival, tie frequency, and competition protocols (Resnick et al., 2018). Alexa Arena reports mission success rate, number of robot actions, COCO mAP, and t-mAP (Gao et al., 2023). MuJoCo Playground reports throughput, zero-shot transfer, task success, and hardware error measures (Zakka et al., 12 Feb 2025). MO-Playground evaluates Pareto-front hypervolume and time to convergence (Janwani et al., 10 Mar 2026). GS-Playground reports rendering throughput, reconstruction quality, contact retention, simulation-to-real success, and simulation–real success correlation (Jia et al., 28 Apr 2026). The Embodied Playground benchmark specifies independent checks of goal completion, task-rule compliance, and physical validity across standardized tasks (Chen et al., 19 Oct 2026).

Despite their diversity, these environments share unresolved limitations:

• Physical safety: running, exertion, public-space play, actuation, and contact-rich manipulation can produce collision, fatigue, or injury risks.
• Accessibility: many systems assume standing, walking, visual tracking, headset use, particular reach, strength, balance, or dexterity.
• Representation bias: a transformed or simulated body may preserve tactical intent while misrepresenting actual mechanics, identity, or capability.
• Technical fragility: tracking failures, latency, calibration, occlusion, rendering artifacts, and incomplete physical models can destabilize agency.
• Evaluation validity: short-term enjoyment, prototype feasibility, or benchmark success does not establish long-term learning, social bonding, clinical benefit, or real-world transfer.
• Simulation gaps: abstract action spaces, pre-cached outcomes, rigid-body assumptions, and limited deformable or contact-rich physics may omit crucial failures.
• Privacy and governance: public MR, biometric systems, digital twins, and human-motion recording raise consent, data ownership, surveillance, and security concerns.
• Reward misspecification: scalar or vector rewards may encode task completion while omitting comfort, naturalness, social acceptability, or safety.
• Facilitator power: live reconfiguration, adaptive nudging, and environment control can support participation but may also become opaque or coercive.
• Open-endedness versus legibility: ambiguity supports exploration and authorship, whereas insufficient feedback causes confusion; explicit guidance improves entry but can narrow interpretation.

Future research is consequently likely to focus on environments that combine physical validity, social coordination, adaptive accessibility, and interpretable evaluation. Relevant directions include dynamic and deformable simulation, real-time digital-twin updating, human-in-the-loop reward and objective learning, context-sensitive safety mechanisms, richer spectator participation, multimodal feedback, user-controllable mappings, embodied policy transfer, and longitudinal studies of learning and retention.

The most general principle is that an Embodied Playground should not be judged solely by visual realism, sensor sophistication, or benchmark performance. Its central criterion is whether it makes relations among body, environment, action, consequence, and other participants available for meaningful exploration. In this sense, the paradigm spans the public arcade, the mixed-reality street, the theatre rehearsal room, the robot simulator, the digital operating room, the inclusive sports visualization, and the long-horizon agent benchmark. Each provides a different answer to the same underlying question: how can an environment become a space where embodied agents learn what they can do, what others can do, and how action transforms the world?