JFM learns a resolution-consistent Jacobian for a rigid-soft finger
Jacobian Flow Matching models tendon-to-joint sensitivity as a continuous field and reduces sampling-dependent prediction error on the authors’ recorded finger trajectories.

Researchers in China released Jacobian Flow Matching (JFM) on October 1, 2026 for modeling a tendon-driven, rigid-soft robotic finger. The method learns actuator-to-joint sensitivity as a continuous field rather than a pointwise mapping tied to one sampling interval. On the authors’ finger, it reduces average single-step RMSE by more than 53% and keeps multi-step predictions more stable when observations become sparse. Paper and version record.
Explore tactile sensing for the broader comparison.
Key takeaways
- The finger combines 3D-printed phalanges, PTFE-coated joint interfaces, a silicone capsule with woven ligaments and a tendon transmission network.
- The collection totals about 10,000 physical samples, split 80/10/10 for training, validation and testing, with hand pose at 45 Hz and servo state at 55 Hz aligned by nearest timestamps.
- The reported results validate prediction and open-loop command recovery; closed-loop ODE control, multi-finger coordination and out-of-plane motion remain future work.
Why a point Jacobian is not enough
Rigid-soft fingers are difficult to model because the same tendon command can produce different motion depending on pose, friction, backlash, hysteresis and viscoelastic state. A discrete learned Jacobian can fit observed endpoints at one controller rate yet behave unpredictably when an optimizer queries intermediate states or a slower sensor produces larger steps.
JFM normalizes each observed transition into a unit-time flow. Conditional Flow Matching supplies intermediate states between the start and end poses, and a consistency loss discourages the network from depending mainly on the starting anchor. The learned field can then serve either a single pointwise update or an ODE rollout that subdivides a command. The paper compares the training objectives with matched TinyTransformer and LSTM backbones and evaluates pointwise and ODE inference separately. Architecture and physical data pipeline.
What the numbers establish
For single-step servo-to-angle prediction, JFM reduces global average RMSE by 57.27% with the TinyTransformer and 53.68% with the LSTM. Under the 0.015 RMSE threshold, 96.2% of TinyTransformer JFM samples qualify versus 77.9% for its baseline; the LSTM comparison is 96.1% versus 78.8%. These are distribution-level results across three joints, not a task success rate.
Resolution tests use 160 held-out trajectories. At stride 1, pointwise and integrated predictions are close. At stride 8, ODE inference reduces median RMSE by 14.43% and variance by 24.87% relative to the pointwise mode. This supports the specific claim that the field remains useful when a transition is subdivided or samples are skipped.
An open-loop inverse test optimizes tendon commands for target joint motion. It reports overall RMSE 1.4151, mean angular error −0.2576° and standard deviation 1.39°. The model smooths abrupt commands to zero at hard stops because those boundary conditions are not explicit in its continuous field. Pointwise inference takes about 3.5 ms per query on the reported workstation, but the paper does not report a closed-loop ODE controller running on hardware.
RoboSkin analysis
The paper addresses a quiet integration problem for robot hands: perception and control rarely run at identical, perfectly stable rates. A model that changes behavior when the update interval changes can make calibration results misleading. JFM’s value is therefore less about another predictor score and more about a testable consistency condition across controller resolutions.
The recorded signals are proprioceptive, including glove-based finger orientation and servo state, not tactile measurements. Contact, friction and material effects are absorbed as disturbances within the training distribution rather than measured explicitly. For teams combining compliant mechanisms with robot-hand tactile sensors, the next question is whether direct contact observations improve the field or reveal regimes where a local first-order model breaks down.
Limitations and availability
Evidence comes from one finger, flexion motion and one hardware/data pipeline. The method assumes mechanical characteristics change slowly within a trial. It has not been directly validated on pneumatic hands, silicone hands, multi-finger coordination or out-of-plane motion. Sparse-sampling results are open-loop prediction tests, and the inverse test does not demonstrate closed-loop task completion.
The manuscript is an arXiv v1 preprint under review at Robotics and Autonomous Systems. It describes ROS 2 bags converted to HDF5, but no public dataset URL, code repository, model checkpoint, CAD package or implementation license was verified. The arXiv page itself does not grant rights to an absent implementation.
Separate interpolation, inference timing and physical validation
JFM normalizes an observed transition into a unit interval. Its intermediate training states and integrated queries are model constructions, not newly measured high-rate finger motion or an independently calibrated instantaneous physical Jacobian. Likewise, evaluating strides of 1, 4 and 8 on held-out recorded trajectories tests sampling consistency; it does not test arbitrary communication delays or a replacement mechanism. Method and evaluation details.
Three checks matter before incorporating this field into a controller:
- Data independence: identify the split unit and keep complete capture sequences, sessions or replacement-hardware trials together as appropriate. The reported sample split alone does not establish generalization to a new day, person or finger.
- Timing and units: inspect nearest-timestamp pairing between the glove and servo streams; record query mode, solver work and the age of the observations. A pointwise query time excludes the rest of a physical control loop.
- Executed motion: distinguish reconstructing recorded tendon commands from applying them and measuring tracking, hard-stop behavior and contact under a new load. The inverse test reported here supplies the former, not closed-loop task completion.
These are proposed follow-up checks. Retain separate angular-error and physical-task criteria, and repeat calibration when tendon tension or contact mechanics changes. A useful motion model can support gripper integration, but does not itself supply contact sensing, a force limit or grasp reliability.
Sources and related resources
- Rigid-Soft Finger JFM v1, submitted October 1, 2026
- Full method, data details and evaluation
- RoboSkin tactile sensor calibration guide
- RoboSkin robot manipulation hub
*Source review updated November 4, 2026. Original publication dates are retained. Experiments are attributed to their authors; RoboSkin did not reproduce the hardware, execute the models or inspect downloaded experimental payloads.*


