Rickmer Krohn

Accepted at IEEE Robotics and Automation Letters 🎉 With @FatAndFurious42 @GabrieleTiboni1 @GeorgiaChal 🌐 Website + videos: msdp-pearl.github.io

⚡ Full pipeline (data + pretraining + online RL) in <55 min 🤏 Only 6,000 real-robot interactions needed 📈 FT-sensor boosts success rate by +14% 🦾 ~90% success on challenging contact-rich tasks Robust to: sensor noise · variable stiffness · external forces · 🪩 disco lights

Solving contact-rich manipulation tasks even under disco light🪩? Incorporate multiple sensors using MSDP and obtain a robust policy in under 55 min! Excited to share our work accepted at RA-L: Self-Supervised Multisensory Pretraining for Contact-Rich Robot Reinforcement Learning

Multisensory Dynamic Pretraining (MSDP) is based on masked autoencoding and cross-sensor prediction, leading to rich sensor fusion. Usage of the FT-sensor boosts performance by 14%! Our approach is flexible regarding the number and type of sensors. More Details in the Preprint!

Real-world online Reinforcement Learning from Vision + ForceTorque + Proprioception --> With MSDP we achieve ~90% success rate in only 35min of online training! Checkout the Preprint here: arxiv.org/abs/2511.14427 Thanks to @FatAndFurious42 @GabrieleTiboni1 @GeorgiaChal!