Learning contactfrom motor current

In imitation learning, a person demonstrates a task and a policy learns the relationship between what it observes in the environment and how it should act. The promise of learning from demonstrations is easiest to realise when the task can be inferred from geometry. In free space, a camera shows the location of objects and joint positions show the position of the arm in space. Contact-rich manipulation is harder because the decisive state may be physical and not purely geometric. A gripper can close without securing an object, an object can slip midway, a peg can reach a hole while pressing against its rim, and two nearly identical camera images can precede either a smooth insertion or a jam. In each case, the robot must respond to an interaction that is only partly visible.

Visual occlusion

This loss of direct visual evidence near contact is a central observability problem in contact-rich imitation learning.

1. 1

   Images and joint positions provide strong information about appearance, configuration, and motion, but they do not directly measure interaction effort.

2. 2

   Contact-aware policies augment vision and kinematics with tactile, force, or torque measurements, but these usually require additional hardware, calibration, or a robot that already provides reliable torque estimates.

3. 3

   The Waveshare ST3215 control table specifies a read-only motor-current register. Reading an existing current register requires no additional sensing hardware.

4. 4

   Its interpretation is difficult because motor-side current is affected by pose, motion, friction, and the high-ratio transmission, and it does not directly specify the direction or location of an external force.

These questions separate three claims that could otherwise be confused: the signal contains task-relevant information, a trained model uses that information, and its use improves closed-loop behaviour.

Characterisation of the current signal

Before motor current can be added to a policy, it must first be shown to carry a repeatable and task-relevant signal.

Dataset

The policies must learn from the same demonstrations, images, joint positions, and action targets, while the additional current input is introduced through a reproducible transformation.

Evaluation

Binary task success is the primary endpoint. A trial succeeds when the peg is fully seated and flush with the fixture surface before the timeout. Every other outcome is a failure.

Results

The strongest favourable task result is obtained with joint-selective temporal ACT. It succeeds in 12/30 trials against 7/30 for its contemporaneous vision-and-position control, a 71.4% relative improvement and a gain of 16.7 percentage points.

The favourable temporal ACT result is of practically relevant observed size, although its difference interval includes zero and does not establish superiority. Since visual-baseline success varies across sessions, it does not establish a controlled ranking of the current representations.

Conclusion

Existing motor-current readings on the evaluated SO-101 contain contact-related information and can influence learned action predictions. The favourable temporal ACT observation supports further investigation of the current-aware setting, while its uncertainty leaves superiority unresolved.

Citation

@mastersthesis{ranpariya2026contact,
  author = {Jaydeepkumar Ranpariya},
  title  = {Learning Contact Representations from Zero-Cost
            Proprioceptive Effort Signals for Contact-Rich
            Visuomotor Imitation Learning},
  school = {Friedrich-Alexander-Universit{\"a}t Erlangen-N{\"u}rnberg},
  year   = {2026},
  note   = {In cooperation with Fraunhofer IIS},
  url    = {https://contact-from-current.pages.dev/}
}