Fan Shi

Assistant Professor in robotics & AI @NUSingapore
Previously @ETH @JSK @PKU1898

[Fan Shi @fanshi_robot]

Reward models shouldn’t just predict progress—they should know when to trust their predictions.

We introduce RARM, a confidence-gated reward model that mitigates over-rewarding and reward hacking, outperforming baselines on 4 real-world tasks and 9 simulation experiments. 🚀

More details 👇
https://t.co/I4VO0o0c8K

[Pengzhi Yang @pengzhi_yang_98]

Illustrated in this LIBERO-Long example, RARM assigns high cumulative rewards to the successful rollout while avoiding over-rewarding the failed one.

Quantitative results further show that, when learning from scratch, RARM consistently outperforms baseline reference-based and VLM-based reward models across diverse tasks; while ablations demonstrate the effectiveness of the mechanisms in RARM.

🧵(4/n)

[Pengzhi Yang @pengzhi_yang_98]

Progress-based reward models have been proposed to deal with reward design for RL in manipulation. 📈

❌ Existing reward models often suffer from over-estimation/over-rewarding, which will lead to reward hacking problems.
⚡️ This highlights the need for reward models that not only predict task progress but also estimate how much to trust it.
🚀 We're introducing RARM, a reward model for training robotic manipulation policies with confidence gating.

🧵(1/n)

[Fan Shi @fanshi_robot]

@junja941 Thanks the RL legend! Synthetic data is still cheapest way to get the high-quality data 🚀

[Fan Shi @fanshi_robot]

3/3 More work on simulation for robot learning: https://t.co/iwpV7FsqUQ

Huge credit to the team for making this happen — @_Siyuan_Luo @ChongZzZhang @bingyang_zzz and many amazing colleagues!

[Fan Shi @fanshi_robot]

MPPI, teacher-student learning, and VLA — all with 100% simulation data, easy, cheap, scalable. 🚀

Deformable–rigid interaction has long been a bottleneck for fast & accurate simulators. Our GPU-native synthetic data pipeline breaks it for deformable sim2real. 🧵

[Fan Shi @fanshi_robot]

2/ Teacher-student training from scratch.

A blank policy learns purely from synthetic data — no expensive teleop data.

50 minutes on one RTX 5090 for towel folding. $0.20 total training cost.

More details: https://t.co/IJhRmaqlKo

[Fan Shi @fanshi_robot]

More details: https://t.co/ttbjV8BE0G

[Fan Shi @fanshi_robot]

𝐇𝐚𝐩𝐩𝐞𝐧𝐢𝐧𝐠 𝐭𝐨𝐦𝐨𝐫𝐫𝐨𝐰 𝐢𝐧 𝐒𝐭𝐫𝐚𝐮𝐬𝐬 𝟑! #ICRA2026

Join us for an exciting program featuring invited talks and poster presentations showcasing state-of-the-art research on 𝐬𝐲𝐧𝐭𝐡𝐞𝐭𝐢𝐜 𝐝𝐚𝐭𝐚 𝐟𝐨𝐫 𝐫𝐨𝐛𝐨𝐭 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠!

[C. Zhang @ChongZzZhang]

Nice one!

We also worked on cloth folding sim2real last year with a custom solver for massive parallelization. The results might look less decent but it is teleop-free.
Release soon!

[Fan Shi @fanshi_robot]

🚨 Call for Posters: ICRA 2026 Workshop on Synthetic Data for Robot Learning

Synthetic data is transforming robot learning, from physics simulation to world models & generative methods.

📢 Submit your novel ideas!
🗓 Apr 3–17
🏆 Best Poster Award
🔗 https://t.co/ttbjV8CbQe

[Fan Shi @fanshi_robot]

Excited to be involved one of them. 2026 will open many new possibilities

[C. Zhang @ChongZzZhang]

Just witnessed some great things that will change the landscape.

Happy new year.

Might release two big works in January.

[Fan Shi @fanshi_robot]

@arthurallshire Congratulations Arthur!

[Fan Shi @fanshi_robot]

@RoboReading I am a beginner in legged robots, do you think that’s a good paper?

[Fan Shi @fanshi_robot]

Welcome to join the #ICRA2025 workshop "𝐓𝐨𝐰𝐚𝐫𝐝𝐬 𝐑𝐞𝐥𝐢𝐚𝐛𝐥𝐞 & 𝐓𝐫𝐮𝐬𝐭𝐰𝐨𝐫𝐭𝐡𝐲 𝐄𝐦𝐛𝐨𝐝𝐢𝐞𝐝 𝐀𝐈 𝐢𝐧 𝐄𝐯𝐞𝐫𝐲𝐝𝐚𝐲 𝐒𝐜𝐞𝐧𝐚𝐫𝐢𝐨𝐬" happened today at 𝐂𝐨𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐑𝐨𝐨𝐦 311 with exciting talks!

Schedule: https://t.co/jT9SnqmsJg

[Fan Shi @fanshi_robot]

@eastskykang It finishes like a proud dog :) Amazing motion on such tough gait 🔥

[Fan Shi @fanshi_robot]

@robot_in_space2 Congratulations! 🚀

[Fan Shi @fanshi_robot]

@YifanHou2 indeed

[Fan Shi @fanshi_robot]

@robot_in_space2 Reduce noise and green power 🍏

[Fan Shi @fanshi_robot]

Excited to be part of this project to reduce noise in Sony Aibo robot, making them more suitable for human-centered deployment!

We also believe industry and academia will collaborate more closely to solve real-world challenges.

Again, kudos to Watanabe-san!

[Robotic Systems Lab @leggedrobotics]

Check out our #ICRA2025 paper where we train a home robot to walk quietly!

Project site: https://t.co/ZUIo0VMeo3
Video: https://t.co/LNErGayl81

Authors: Ryo Watanabe, Takahiro Miki, Fan Shi, Yuki Kadokawa, Filip Bjelonic,
Kento Kawaharazuka, Andrei Cramariuc and Marco Hutter

[Fan Shi @fanshi_robot]

@Xingyu2017 Thanks Xingyu! Definitely deserves to do