The ability to conduct and learn from interaction and experience is a central challenge in robotics, offering a scalable alternative to labor-intensive human demonstrations. However, realizing such "play" requires (1) a policy robust to diverse, potentially out-of-distribution environment states, and (2) a procedure that continuously produces useful robot experience. To address these challenges, we introduce Tether, a method for autonomous functional play involving structured, task-directed interactions. First, we design a novel open-loop policy that warps actions from a small set of source demonstrations (<=10) by anchoring them to semantic keypoint correspondences in the target scene. We show that this design is extremely data-efficient and robust even under significant spatial and semantic variations. Second, we deploy this policy for autonomous functional play in the real world via a continuous cycle of task selection, execution, evaluation, and improvement, guided by the visual understanding capabilities of vision-language models. This procedure generates diverse, high-quality datasets with minimal human intervention. In a household-like multi-object setup, our method is the first to perform many hours of autonomous multi-task play in the real world starting from only a handful of demonstrations. This produces a stream of data that consistently improves the performance of closed-loop imitation policies over time, ultimately yielding over 1000 expert-level trajectories and training policies competitive with those learned from human-collected demonstrations.
Figures & tables
Figure 1: Tether performs autonomous functional play in the real-world for over 24 hours, streaming over 1000 successful trajectories for downstream policy learning.
Figure 2: Demonstration Summaries. Tether summarizes demonstrations into the initial frame, action sequence (red), waypoints (blue), and keypoints.
Figure 3: Policy Inference . During inference, Tether (left) computes correspondences (middle) and produces a warped trajectory action plan (right).
Figure 4: Autonomous Functional Play. Our iterative procedure runs Tether for multiple tasks and uses VLMs for plan generation and success detection.
Figure 5: Evaluation Tasks. Our tasks involving moving fruits and containers with in-distribution (orange) and out-of-distribution (green) objects, as well as challenging manipulation skills (purple).
Figure 6: Main Policy Comparison. We compare the Tether policy with baselines across 12 tasks.
Figure 7: Autonomous Play Statistics. In around 26 hours of play, our method produces over 1000 cumulative trajectories across 6 tasks (left) and significantly expands data diversity, as shown by the heatmaps of object poses at the beginning and end of play (right).
Figure 8: Downstream Policy Learning Results. The stream of data generated by autonomous play consistently improves policy performance over time. Policies trained on human-collected demos, equal in number to the final set of successful Tether-generated trajectories, are shown in black.
Figure 9: Comparison on Play Distribution. We compare Diffusion Policies trained on human-collected demos with Tether policy on environment states encountered during autonomous play.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Pineapple from Shelf
Pineapple to Bowl
Bowl to Shelf
Bowl from Shelf
Tether (10 Demos)
1.0 ± 0.0
1.0 ± 0.0
0.9 ± 0.09
1.0 ± 0.0
Tether (5 Demos)
0.8 ± 0.13
0.8 ± 0.13
1.0 ± 0.0
1.0 ± 0.0
Tether (1 Demo)
0.8 ± 0.13
0.9 ± 0.09
0.7 ± 0.14
1.0 ± 0.0
π0 ( Black et al., 2024 )
0.5 ± 0.16
0.7 ± 0.14
0.0 ± 0.0
0.4 ± 0.15
DP ( Chi et al., 2023 )
0.0 ± 0.0
0.0 ± 0.0
0.0 ± 0.0
0.1 ± 0.09
Method
Apple to Bowl
Strawberry to Bowl
Pineapple to Basket
Pineapple to Cup
Appendix
Table 1: Success rates and standard error measured over 10 trials.
Figure 10: Computed correspondence for the Pineapple to Cup task.
Figure 11: Computed correspondence for the Strawberry to Bowl task.
Figure 12: Computed correspondence for the Hanging Tape task.
Figure 13: Computed correspondence for the Wiping with Cloth task.
Figure 14: Trajectory warping for the Pineapple to Bowl task.
Figure 15: Trajectory warping for the Wiping with Cloth task.
Figure 16: Environment states with the flipped bowl (left) or pineapple pushed into shelf (right) that required manual interventions.
Method
Pineapple from Shelf
Pineapple to Bowl
Bowl to Shelf
Bowl from Shelf
Tether
1.0 ± 0.0
1.0 ± 0.0
0.9 ± 0.09
1.0 ± 0.0
KALM
0.6 ± 0.15
0.0 ± 0.0
0.0 ± 0.0
0.5 ± 0.16
Appendix
Table 2: Policy Comparison with KALM. Success rates and standard error measured over 10 trials.
Method
Pineapple from Shelf
Pineapple to Bowl
Bowl to Shelf
π0 (10 human demos)
0.0 ± 0.0
0.0 ± 0.0
0.0 ± 0.0
π0 (50 play-generated demos)
0.1 ± 0.09
0.2 ± 0.13
0.1 ± 0.09
π0 (100 play-generated demos)
0.3 ± 0.14
0.5 ± 0.16
0.1 ± 0.09
π0 (all play-generated demos)
0.7 ± 0.14
0.8 ± 0.13
0.4 ± 0.15
π0 (pre-trained)
0.5 ± 0.16
0.7 ± 0.14
0.0 ± 0.0
Appendix
Table 3: Downstream π0 Finetuning Results. Success rates and standard error measured over 10 trials.
Action-conditioned video models offer a promising path to building general-purpose robot simulators that can improve directly from data. Yet, despite training on large-scale robot datasets, current state-of-the-art video models still struggle to predict physically consistent robot-object interactions that are crucial in robotic manipulation. To close this gap, we present PlayWorld, a simple, scalable, and fully autonomous pipeline for training high-fidelity video world simulators from interaction experience. In contrast to prior approaches that rely on success-biased human demonstrations, PlayWorld is the first system capable of learning entirely from unsupervised robot self-play, enabling naturally scalable data collection while capturing complex, long-tailed physical interactions essential for modeling realistic object dynamics. Experiments across diverse manipulation tasks show that PlayWorld generates high-quality, physically consistent predictions for contact-rich interactions that are not captured by world models trained on human-collected data. We further demonstrate the versatility of PlayWorld in enabling fine-grained failure prediction and policy evaluation, with up to 40% improvements over human-collected data. Finally, we demonstrate how PlayWorld enables reinforcement learning in the world model, improving policy performance by 65% in success rates when deployed in the real world.
Current recipes for robot learning require massive amounts of robot demonstrations to obtain robust and adaptive policies. In this work, we introduce Steertuning, a system that enables robots to learn new tasks in new environments from 5 human demonstrations in under 5 Steertuning consists of two parts. (1) We pretrain a 400M-parameter diffusion policy on our egocentric human dataset, which consists of 100 hours of densely annotated random interactions. This pretraining makes Steertuning robust to background differences and enables spatial and object generalization when deployed on new tasks. (2) Given 5 demonstrations of a new task, we finetune our model for only 2 minutes (2000 training updates) and steer the policy's actions during reverse sampling using the demonstration whose observation is closest to that of the deployment environment. Together with a carefully designed data processing pipeline for embodiment transfer, this enables Steertuning to achieve a 71% success rate across 12 tasks in 12 new environments, ranging from lab settings to kitchens in new homes. The entire process, from data collection for a new task to robot policy deployment, takes less than 5 minutes. Our full dataset, model weights, codebase, and robot videos will be made available at our website: https://steertuning.github.io.
Self-play reinforcement learning has recently emerged as a way to train driving policies without any human data. It uses cheap, large-scale simulations to substitute expensive, large-scale human driving demonstrations. A key limitation of this approach is that policies trained through pure self-play can learn effective but alien driving conventions incompatible with people. Previous works attempt to mitigate such behavioral misalignments through extensive reward engineering and domain randomization, which are brittle and labor-intensive. Instead of completely discarding human demonstrations, our method treats them as a regularization objective on top of a minimal safe goal-reaching reward. Like the spice in a good stew, we find that a little human data goes a long way: our method uses only 30 minutes of human demonstrations, 2500x fewer than comparable imitation learning approaches. Resulting policies coordinate with held-out human trajectories and complete training in 15 hours on a single consumer-grade GPU. Videos and full source code are available at https://spiced-self-play.com/.
Daphne Cornelisse, Julian Hunt, Zixu Zhang +4
NYU Tandon School of Engineering · NYU Courant · Princeton University +2