Peking University's TacMan-Turbo Opens Any Drawer 33x Faster With Perfect Accuracy
A proactive tactile controller keeps 100% success on 200 articulated objects while cutting task time roughly tenfold versus the prior tactile baseline.
ProTakeaways
- TacMan-Turbo manipulates articulated objects with 100% success and no kinematic prior, roughly 10x faster than prior tactile methods.
- Key idea: read tactile contact deviations as time-series kinematic data, not spatial errors.
- Evaluated on 200 simulated objects spanning prismatic, revolute, and Bézier trajectories.
- Mean task time drops from 250.7 s to 7.6 s versus Tac-Man, with p < 0.0001 on all metrics.
- Runs on Kinova Gen3 with GelSight sensors at 3.7 ms per control step, training-free.
- Works zero-shot on charging docks, power switches, coffee makers, and sewing machines.
Opening drawers, doors, microwaves, and any object with a joint is one of those things robots are expected to do the moment they leave the lab, and one of the things they still struggle with. A team from Peking University has released TacMan-Turbo, a training-free tactile controller that manipulates articulated objects without any prior kinematic model and does it about an order of magnitude faster than the previous state-of-the-art tactile method, while keeping a perfect success rate.
The core idea: analyze sequential tactile observations as time-series data that reveal local kinematic information, rather than treating contact deviations as spatial errors requiring immediate compensation. That shift from spatial correction to temporal prediction is what unlocks the speedup.
The trade-off nobody could break
Articulated object manipulation has been stuck between two bad options. Planners that assume a known joint model (revolute, prismatic, with fixed parameters) run smoothly when the model is right and fall apart when it isn't. Learning-based approaches like FlowBot3D generalize poorly outside their training distribution. The recent tactile-informed baseline, Tac-Man, achieves rock-solid manipulation without any prior kinematic knowledge, but its reactive execute-and-recover cycle carries significant time overhead.
Concretely, Tac-Man works in two alternating stages. It moves the gripper a bit, detects the deviation between the contact patch it sees now and the reference patch it saw at the start, then stops and executes a corrective motion to bring the contact back into alignment. Repeat until the drawer is open. The result is safe, robust, and painfully slow, with visible zig-zag trajectories.

Treating tactile data as time series
TacMan-Turbo keeps Tac-Man's contact-based state representation but reinterprets what those contact deviations mean. Rather than viewing each deviation as an error to null out, the controller looks at a short history of deviations and treats them as measurements of the handle's motion. From that it predicts where the handle will be one step in the future, then commands the gripper to be there proactively.