RoboCade: Gamifying Robot Data Collection (2512.21235v2)

Abstract: Imitation learning from human demonstrations has become a dominant approach for training autonomous robot policies. However, collecting demonstration datasets is costly: it often requires access to robots and needs sustained effort in a tedious, long process. These factors limit the scale of data available for training policies. We aim to address this scalability challenge by involving a broader audience in a gamified data collection experience that is both accessible and motivating. Specifically, we develop a gamified remote teleoperation platform, RoboCade, to engage general users in collecting data that is beneficial for downstream policy training. To do this, we embed gamification strategies into the design of the system interface and data collection tasks. In the system interface, we include components such as visual feedback, sound effects, goal visualizations, progress bars, leaderboards, and badges. We additionally propose principles for constructing gamified tasks that have overlapping structure with useful downstream target tasks. We instantiate RoboCade on three manipulation tasks -- including spatial arrangement, scanning, and insertion. To illustrate the viability of gamified robot data collection, we collect a demonstration dataset through our platform, and show that co-training robot policies with this data can improve success rate on non-gamified target tasks (+16-56%). Further, we conduct a user study to validate that novice users find the gamified platform significantly more enjoyable than a standard non-gamified platform (+24%). These results highlight the promise of gamified data collection as a scalable, accessible, and engaging method for collecting demonstration data.

• The paper introduces RoboCade, a gamified framework that transforms robot demonstration collection into an engaging, remote teleoperation experience.
• It shows that co-training with gamified data significantly improves policy performance and generalization, with up to 48% boost in out-of-distribution cases.
• User studies confirm that RoboCade enhances intuitiveness, engagement, and system usability compared to traditional teleoperation methods.

RoboCade: Gamifying Scalable Robot Data Collection

Motivation and Problem Setting

Contemporary imitation learning approaches for robotic manipulation are bottlenecked by the cost and scale of collecting real-world expert demonstrations. Traditional paradigms require dedicated, in-person teleoperation by skilled operators, limiting data diversity and throughput. Although remote teleoperation solutions have emerged, they seldom emphasize user engagement, resulting in suboptimal participation and limited scaling. "RoboCade: Gamifying Robot Data Collection" (2512.21235) introduces a novel gamified platform, RoboCade, to address both accessibility and incentive structures, reframing robot data collection as an intrinsically rewarding, interactive experience for a broad user base.

Figure 1: RoboCade contrasts conventional in-person teleoperation with an accessible, gamified remote platform, democratizing robot demonstration collection.

System and Task Design Principles

RoboCade leverages a low-cost, 3D-printable GELLO controller, enabling intuitive remote teleoperation of a Franka FR3 manipulator. The web-based interface integrates gamification elements—visual feedback, goal visualization, progress bars, leaderboards, badges, and auditory rewards—intended to cultivate both intrinsic motivation and social engagement. Crucially, the interface design is informed by four principles: immediate feedback and reward, explicit challenge and goal cues, identity progression, and social competition. The architecture is purpose-built for low-latency data streaming and ease of deployment, supporting robust remote teleoperation across wide geographic regions.

Figure 2: System architecture fuses remote robot control with gamification features to drive engagement and scalability.

On the task level, RoboCade introduces a taxonomy of manipulation challenges organized as pairs of target and support tasks. Support tasks are constructed via narrative overlay, objective diversity, calibrated challenge, and intentional alignment of underlying manipulation skills with downstream tasks. This framework transforms functional manipulation goals into engaging scenarios with thematic and structural variety while preserving the utility of the collected trajectories for policy training.

Figure 3: Task design maps manipulation goals to engaging support tasks, ensuring overlap in skill requirements and enhancing engagement.

Experimental Evaluation and Numerical Results

The core experimental question is whether gamified data collected via RoboCade can accelerate or improve policy learning for traditional manipulation tasks. Demonstrations were collected for three task classes: spatial arrangement (ArrangeDesk/SceneTwins), scanning (ScanBottle/GroceryCheckout), and insertion (PackBox/AnimalDorms). Using both Diffusion Policy [chi2023diffusion] and fine-tuned vision-language-action (VLA) models ($\pi_{0.5}$ [black2025pi_05]), comparative policy learning experiments were performed:

• Diffusion Policy Co-training: Policies trained solely on target data underperformed relative to those co-trained on support tasks. Noteworthy improvements were observed: ArrangeDesk success rate increased from 12% (target-only) to 28% (co-train); ScanBottle from 32% to >50% for in-distribution, with >20% boost in OOD generalization; PackBox exhibited a 16% improvement with superior grasp alignment and placement accuracy.

  Figure 4: Diffusion Policy co-training with gamified data systematically exceeds target-only training across all evaluation protocols.

• VLA Co-fine-tuning: Starting from $\pi_{0.5}$ fine-tuned on the DROID dataset, further co-fine-tuning with RoboCade support data matched or surpassed single-task fine-tuning in-distribution (e.g., ScanBottle +12%), while markedly improving OOD generalization (ScanBottle +48%, PackBox +36%). Enhanced stability and precision in execution were qualitative outcomes.

  Figure 5: VLA model co-fine-tuned with gamified support data achieves higher OOD robustness and maintains strong in-distribution performance.

User Engagement Analysis

A controlled user study (N=18) with novice participants, balanced across interface conditions (gamified vs. vanilla remote teleoperation), measured perceived intuitiveness, enjoyment, and motivation on standard Likert scales. RoboCade was found to be significantly more intuitive (+27%), enjoyable (+24%), and motivating (+24%). Completion rates and System Usability Scale (SUS) scores favored RoboCade (completion: 77.1% vs. 69.7%, SUS: $71.8 \pm 20.4$ vs. $51.4 \pm 22.4$).

Figure 6: User study demonstrates substantial advantages of RoboCade in intuitiveness, enjoyment, and motivational impact for novice operators.

Practical Interface Details

RoboCade’s frontend is implemented with Next.js and Three.js, incorporating live 3D point cloud rendering, dynamic camera movement, and real-time effects. The interface supports persistent user identity, point-based progression, and social features such as leaderboard rankings and shareable session URLs. Backend communication is maintained through ZeroMQ and WebSockets; user authentication via Supabase. The system architecture is optimized for safety (bonded joint limits, collision avoidance) and efficient data streaming in mixed teleoperation scenarios.

Figure 7: RoboCade’s interface design enables remote login, adaptive task selection, immersive gameplay, and post-task progression feedback.

Implications and Future Directions

RoboCade establishes gamification as a practical scaffold for scaling robot demonstration collection without relying on extrinsic incentives. Co-training with gamified support data leads to robust improvements in policy generalization and performance, even within large pre-trained VLA model frameworks. User engagement metrics substantiate the platform’s potential to democratize participation and unlock new pools of data contributors, particularly among those motivated by gameplay rather than academic or technical objectives.

The work also opens directions for dynamic quality control (e.g., through data-driven point rewards), long-term retention studies, and synthesis of play-driven, task-agnostic trajectories to support general robot learning paradigms (e.g., world models, goal-conditioned RL). The intersection of gamification and robotics offers a rich avenue for augmenting both the quantity and diversity of robot training data and for integrating learning pipelines with human creativity and agency.

Conclusion

RoboCade demonstrates the viability and utility of gamifying robot data collection at both the systems and task design levels. Empirical results indicate that data sourced from interactive, narrative-driven manipulation tasks can significantly improve policy efficacy and generalization, particularly when used as part of co-training pipelines. The platform’s strong user engagement and scalability characteristics, combined with robust policy performance gains, highlight gamification as a key lever for future large-scale, sustainable robot learning infrastructure.