OmniTacTune adds tactile layer, lifts robot success to 85-100%

Contact-rich manipulation is the part of the field where the pretty visual policies stop working. A pretrained model can watch a lot of teleop and human video and still fumble tasks that need force or contact geometry awareness, because vision alone does not tell it how hard it is pressing or whether the grip actually landed. A new arXiv paper from Kelin Yu, Haode Zhang, Harish Ravichandar, Yunhai Han and Ruohan Gao proposes a modest-sounding fix that reports very immodest numbers.

Their system, called OmniTacTune, treats a pretrained visual policy as a fixed motion prior and adapts a lightweight tactile residual on top of it, trained with real-world reinforcement learning in two stages: a bootstrap phase of autonomous rollouts, followed by online tactile residual learning. The reported result is that success rates on contact-rich tasks that started at 5 to 40 percent with visual-only policies climbed to 85 to 100 percent after 40 to 80 minutes of real-world interaction. The authors say the recipe held across multiple tasks, multiple base visual policies, and diverse tactile sensor representations.

The strategic pitch, if the numbers hold up outside the authors' own benchmarks, is that robotics teams do not need to retrain their base model every time a new sensor shows up on the arm. Frozen vision policy, small tactile head, an hour or so of RL on the actual robot. That is a very different economic story from the dominant one, which asks teams to collect large paired vision-and-touch datasets or fine-tune the whole policy end-to-end.

The honest caveat is that these are the authors' own tasks and their own success bar, and the work is a fresh arXiv preprint rather than a peer-reviewed replication. What the abstract does not spell out is which specific tactile sensors were tested, how big the residual network is at inference, or whether the approach still lifts a visual base policy that starts even weaker than 5 percent. Even with those unknowns, the direction is worth watching. Contact-rich manipulation has been the stubborn ceiling on foundation-model robotics, and a plug-in tactile adapter that trains in under 80 minutes is exactly the kind of unglamorous piece that would let smaller labs and manipulation startups keep pace with the well-funded ones.