Introducing Tether 🪢, a fun little idea to scale data by having our robot “play” in the real world for over 24 hours, throughout the day and overnight—improving policies from zero to mastery with minimal supervision! But play is messy, with out-of-distribution scenarios that are hard to anticipate. To... perform autonomous functional play in the real world, from just a handful of demos, we propose a highly robust few-shot imitation method that warps demo trajectories using visual correspondences. Then, continuously running it within a multi-task VLM-guided cycle, we generate a data stream that produces 1000+ expert-level demos. This generated data is finally funneled downstream to train imitation learning policies, which improve from zero to near-perfect success rates. We’ll be presenting Tether at #ICLR2026 in just a few weeks! But before that, deep dive with me… 🧵show more

So many unforgettable memories with the group at PennPAL! Photos taken the morning after our last all-nighter, and a day before I donned graduation robes. 😁 @johnnywang_16 @JieWang_ZJUI @JunyaoShi @edward_s_hu and Xingfang

🌐 Website: 📄 Paper: 🐙 Code:

Most methods today rely on heavy human involvement, whether for collecting teleoperation data or supervising real-world learning. We instead minimize manual effort—while scaling 10 demos per task to 1000+ cumulatively, and downstream policy success rates from 0% to 80-100%.

To do so, we design an open-loop policy that’s especially robust to OOD initial states, using visual correspondences and trajectory warping. Running it in a VLM-guided multi-task play loop naturally induces environment resets and diversity, without requiring constant supervision.

First, let’s assess the capabilities of our policy design. Given just 10 demos, our Tether policy surpasses baselines on these 12 tasks, generalizing to unseen objects and performing challenging skills with millimeter-level margin of error.

Next, we run autonomous functional play through day and night for 26+ hours, producing a stream of data for the pineapple and bowl, and averaging 1 successful trajectory every 86 seconds. See this partial timelapse (100x speed)!

During play, we only intervened 5 times (0.26% of attempts). Most involve the bowl flipping, which is nearly impossible to fix with one arm. Yet, we observe 2 spontaneous corrections by chance, highlighting that at scale, coincidences may result in unexpected behaviors!

Finally, we use the generated data for downstream closed-loop policy learning. Training diffusion policies periodically as data streams in, we observe consistent improvements over time, reaching ≥90% on most tasks—competitive with policies trained on human-collected demos!

To conclude: we learned a lot, and there’s still much left to do! Tether decouples play from policy learning, and a key next step is to unify them and bring out the best of both worlds: robust exploration in the low-data regime, and consistent policy improvements over time.

This was the last project from my undergrad at Penn ❤️💙, with my team @sam_wang23 @johnnywang_16 and advisors @JasonMa2020 @dineshjayaraman @obastani! I’m so fortunate to work with and learn from these amazing researchers, through late nights and very-early mornings. Cheers!

Cheers! It’s awesome to work with you on Franka infra and hope to see more exciting stuff coming soon

Had a lot of fun working with the amazing team @sam_wang23 and @willjhliang on this project. Still remembers when it “played” through day and night!

are you accepting donations for tether via github?

This experimentation is what pushes robotics forward. What kind of failure modes popped up during overnight play?

Mastery through play is efficient. But the real friction is managing the physical entropy of 24-hour autonomy.