RoboCade: Gamifying Robot Data Collection

TL;DR

The paper addresses the scalability bottleneck in imitation learning by introducing RoboCade, a gamified remote teleoperation platform that engages a broad audience to collect robot demonstrations. It combines system-level gamification (feedback, progression, social elements) with task-level design (gamified support tasks that share skills with downstream target tasks) and validates the approach through policy learning and human-subject studies. Results show that co-training policies with gamified data improves in-distribution success across multiple target tasks and enhances generalization, while users report higher enjoyment and motivation compared with non-gamified teleoperation. Overall, RoboCade demonstrates a scalable, accessible, and engaging approach to collecting high-quality demonstration data on real robots, with tangible gains in policy performance and user engagement.

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.

Figures (7)

• Figure 1: We develop RoboCade, a platform that gamifies the collection of robot demonstration datasets. In standard data collection approaches (top), hired operators or researchers collect demonstrations using in-person teleoperation, which requires access to a robot and can be tedious and time-consuming. RoboCade (bottom) is a remote data collection platform that integrates gamification into both system and task design, making robot data collection more engaging and accessible to a broader set of users.
• Figure 2: System and Task Design Overview. RoboCade uses a GELLO controller wu2024gello with a web-based interface to enable remote teleoperation of real robots. We embed gamification principles into the design of the system and tasks. (Top) Into the system, we incorporate visual and auditory feedback and rewards, challenge and goal visual cues, identity and progression, and social engagement. (Bottom) Into the design of tasks, we incorporate narrative, goal diversity, an appropriate level of challenge, and overlapping skills with relevant downstream tasks.
• Figure 3: Task Design. We show 3 pairs of support tasks (right) and target tasks (left). Rearrangement (top): Both tasks involve moving 2 objects; ArrangeDesk focuses on organizing electronics, while SceneTwins uses animal blocks with diverse layouts and a virtual goal overlay. Scanning (middle): both tasks involve grasping and aligning objects for scanning; ScanBottle requires more precision due to the object geometry and randomized orientations, while GroceryCheckout adds object variety, a grocery checkout narrative, sound effects, and a basket placement step. Insertion (bottom): both tasks require packing, but PackBox adds complexity by closing a box lid while AnimalDorms adds narrative and object variety.
• Figure 4: Co-training with Gamified Data. For our 3 target tasks, we compare the performance of Diffusion Policy chi2023diffusion when trained only on target task data (Target Only) versus co-trained with gamified support tasks (Co-train), with a fixed training budget. For each task and training condition, we perform 25 trials and report staged success rate (%). Co-training improves success rate on all 3 tasks for in-distribution conditions (In-Dist.). For ScanBottle and PackBox, we additionally evaluate on out-of-distribution initial configurations (Out-of-Dist.) and find that co-training improves generalization.
• Figure 5: VLA Co-fine-tuning with Gamified Data. We compare the performance of a version of $\pi_{0.5}$black2025pi_05 fine-tuned on the DROID dataset Khazatsky2024DROIDAL when further fine-tuned on target task data versus co-fine-tuned on both target task and gamified support task data, with a fixed budget of training steps. For each task and training condition, we perform 25 trials. Co-fine-tuning on gamified support task data improves performance in out-of-distribution initial configurations (Out-of-Dist.) while matching or exceeding performance in in-distribution conditions (In-Dist.).