Blur the Camera, Leak the Joints: Private Scene Inference from Robot Proprioceptive and Action Telemetry

Physical-AI companies increasingly treat the data their robots stream home — joint angles, actuator torques, and the action sequences a policy emits — as the strategic fuel of the "data flywheel." Camera frames in these uploads are often redacted for privacy (faces blurred, pixels dropped). We show that this redaction creates a false sense of safety: the low-dimensional proprioceptive and action telemetry that is uploaded as-is, because it is assumed to be innocuous engineering data, leaks private information about the scene, the objects a robot handles, and the people around it. We demonstrate three attacks in high-fidelity physics simulation (MuJoCo; a Franka Panda arm and a mobile home robot):(A) from wrist/joint torque telemetry, an adversary recovers the mass of a grasped object to a median error of 2.7 g and classifies sealed medication packages into sensitive therapeutic categories (e.g. antidepressant, HIV PrEP, antipsychotic) with 94.5% accuracy;(B) from joint-angle telemetry alone, with the camera removed, an adversary recovers a typed keypad PIN with 71.1% exact-sequence and 86.3% per-digit accuracy;(C) from the action stream of a camera-conditioned navigation policy, with images deleted from the logs, an adversary infers whether the robot encountered a person (100% accuracy), the goal room (45.5% of 5 classes vs. 21.9% chance), and dead-reckons the route to recover the floor plan. The signal survives aggressive downsampling and sensor noise because it is rooted in the geometry and dynamics of motion, not in any single sample. We then present TeleDLP, a leakage-aware telemetry-release layer, and chart the privacy–utility trade-off using a concrete utility metric (behaviour cloning from the sanitised logs), so the cost of protection is explicit rather than assumed. TeleDLP's context-aware release reaches the same strong-privacy regime as uniform noise at roughly half the utility cost. Our results argue that robot telemetry content, not just the camera, must be governed as personal data. All experiments are in simulation; we discuss the sim-to-real gap and release no attack code.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24221587
Primary Topic
Privacy-Preserving Technologies in Data
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Blur the Camera, Leak the Joints: Private Scene Inference from Robot Proprioceptive and Action Telemetry

Gorkem Yagci, Cinar Fidanboy, Asil Ozyildirim
Zenodo (CERN European Organization for Nuclear Research)
Privacy-Preserving Technologies in Data
preprint

Blur the Camera, Leak the Joints: Private Scene Inference from Robot Proprioceptive and Action Telemetry

Gorkem Yagci, Cinar Fidanboy, Asil Ozyildirim
preprint en

Abstract

Physical-AI companies increasingly treat the data their robots stream home — joint angles, actuator torques, and the action sequences a policy emits — as the strategic fuel of the "data flywheel." Camera frames in these uploads are often redacted for privacy (faces blurred, pixels dropped). We show that this redaction creates a false sense of safety: the low-dimensional proprioceptive and action telemetry that is uploaded as-is, because it is assumed to be innocuous engineering data, leaks private information about the scene, the objects a robot handles, and the people around it. We demonstrate three attacks in high-fidelity physics simulation (MuJoCo; a Franka Panda arm and a mobile home robot):(A) from wrist/joint torque telemetry, an adversary recovers the mass of a grasped object to a median error of 2.7 g and classifies sealed medication packages into sensitive therapeutic categories (e.g. antidepressant, HIV PrEP, antipsychotic) with 94.5% accuracy;(B) from joint-angle telemetry alone, with the camera removed, an adversary recovers a typed keypad PIN with 71.1% exact-sequence and 86.3% per-digit accuracy;(C) from the action stream of a camera-conditioned navigation policy, with images deleted from the logs, an adversary infers whether the robot encountered a person (100% accuracy), the goal room (45.5% of 5 classes vs. 21.9% chance), and dead-reckons the route to recover the floor plan. The signal survives aggressive downsampling and sensor noise because it is rooted in the geometry and dynamics of motion, not in any single sample. We then present TeleDLP, a leakage-aware telemetry-release layer, and chart the privacy–utility trade-off using a concrete utility metric (behaviour cloning from the sanitised logs), so the cost of protection is explicit rather than assumed. TeleDLP's context-aware release reaches the same strong-privacy regime as uniform noise at roughly half the utility cost. Our results argue that robot telemetry content, not just the camera, must be governed as personal data. All experiments are in simulation; we discuss the sim-to-real gap and release no attack code.

Zenodo (CERN European Organization for Nuclear Research)
Politecnico di Torino (IT), Hospital Privado (AR)
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon