Master's thesis / 2026

Learning Contact
from Motor Current

Learning Contact Representations from Zero-Cost Proprioceptive Effort Signals for Contact-Rich Visuomotor Imitation Learning

Read the study Reproduction record

Platform: SO-101 / STS3215
Task: Peg-in-hole insertion
Institution: FAU / Fraunhofer IIS
Author: Jaydeepkumar Ranpariya

Vertical peg-in-hole demonstration. One human-reviewed demonstration from approach to release, shown in synchronized wrist and side views.

Study

One signal, three questions

Near contact, similar camera images can require different actions. A peg may be aligned with the hole, pressing on its rim, or beginning to jam. Force and tactile sensors can expose these states, but they add hardware and integration work.

The ST3215 servos on the SO-101 already report motor current. The reading costs nothing to add to the observation, but gravity, motion, gearbox friction, grasp load, and contact can all affect it.

01 / SignalDoes current contain task-relevant variation?
02 / ModelDo trained policies respond when current changes?
03 / RobotDoes adding current improve insertion success?

Physical rollouts

What the policies did

Individual recordings make the observed behaviour tangible. They illustrate the aggregate comparisons below and are not paired trials.

ACT vision baselineMissed insertion

The policy reaches insertion but does not seat the peg. ACT-V, block 3, global trial 13

Temporal current ACTSuccessful insertion

A current-aware ACT rollout seats the peg in the same evaluation block and trial slot. ACT-M temporal, block 3, global trial 13

Current-aware Diffusion PolicySuccessful insertion

A rollout from the final Diffusion Policy comparison seats the peg. DP-M learned, block 1, global trial 2

01 / Signal characterisation

Structured variation, not a force measurement

The six servo-current channels were examined under static poses, known loads, scripted grasps, natural motion, and reviewed insertion phases.

Baseline-corrected joint-current changes during 30 reviewed insertion phases — Baseline-corrected current changes during 30 reviewed insertion phases. Several joints carry coordinated, task-aligned variation above the quiet-state noise floor.

Dataset

A synchronized multimodal record

One hundred vertical peg-in-hole demonstrations were collected through leader-follower teleoperation.

Wrist and side views of approach, grasp, transport, insertion, and release — Approach, grasp, transport, insertion, and release in one reviewed demonstration. The principal policies receive the wrist view; the side view supports inspection and annotation.

Demonstrations: 100
Synchronized frames: 99,831
Nominal frame rate: 30 Hz
Image streams: 2
Current channels: 6
Controlled data views: 3

The policy views retain identical episodes, frame indices, images, timestamps, and action targets. Only the state input changes. Current caps and normalization statistics are fitted on training demonstrations alone.

03 / Policy inputs

Four ways to expose current

The controlled comparisons keep episodes, frames, images, timestamps, and action targets fixed. Only the state view changes.

ACT-V

Vision and position

A wrist image and six follower joint positions provide the reference input.

ACT direct

One current sample per joint

Six instantaneous current values are scaled with training-derived statistics and appended to the state vector.

ACT temporal

Ten causal samples per joint

A shared one-dimensional convolutional encoder maps each current history to a joint-specific representation.

Diffusion Policy

Learned residual conditioning

A temporal current route conditions a frozen vision and position base through bounded residual modulation.

04 / Model response

Does current reach the output?

Current is replaced with its training mean while image and joint position remain fixed. The action change measures checkpoint sensitivity to the added modality.

Action changes after current, position, and image interventions — Normalized action change after each intervention. These perturbations are not physically equivalent, so the values establish model response rather than causal importance.

05 / Closed-loop evaluation

Four completed comparisons

Success requires the peg to be fully seated and flush before timeout. Each effect is the current-aware success rate minus its contemporaneous vision-only rate.

Differences in task success for four current-aware policy comparisons with 95 percent confidence intervals — Differences in task success with Newcombe-Wilson 95% intervals. The raw direct ACT interval excludes zero in the harmful direction; the other intervals include zero.

Comparison	Vision only	Current aware	Difference, percentage points (95% CI)
Raw direct ACT	19/30	8/30	−36.7 [−56.0, −11.5]
Normalized direct ACT, two sessions	27/50	21/50	−12.0 [−30.1, 7.4]
Learned temporal ACT	7/30	12/30	+16.7 [−6.7, 37.8]
Learned residual Diffusion Policy	9/30	12/30	+10.0 [−13.6, 32.1]

Success-rate variation across ACT sessions and Diffusion Policy evaluation blocks — The unchanged ACT-V checkpoint varies across sessions, while combined Diffusion Policy success declines across blocks from 8/12 to 1/12. Within-session controls are therefore essential.

Trial progress from evaluation start to insertion and success — Most failures of the later policies occur after reaching insertion, but the pattern does not repeat across policy families strongly enough to identify a phase-specific benefit from current.

06 / Conclusion

Information, model response, and control benefit are different claims

STS3215 current contains contact-relevant information on the evaluated SO-101, and several retained policies respond when that input is changed. Under the evaluated data, fusion, and execution settings, neither ACT nor Diffusion Policy demonstrates a reliable improvement in peg-in-hole success from adding current.

The study’s central contribution is to evaluate those claims separately through signal characterization, controlled policy comparisons, counterfactual diagnostics, and repeated physical trials.

Project record

Evidence behind the published numbers

The reproduction record maps the numerical claims on this site to retained data, model identities, trial records, and deterministic checks. The reviewed thesis PDF and public archive will follow after submission.

Open the reproduction record