SPINE

Scalable Physical Integration with ageNtic Expertise

Bridging the Cyber‑Physical Gap
with Agentic AI

Minkyu Ham1,*, Dongho Kim1,*, Chan Lee1,*, Min Jun Kim1, Yixi Zhang2, Jiayi Wang2, Han Liu1,2,†

1Department of Statistics and Data Science 2Department of Computer Science
Northwestern University

*Equal contribution, alphabetical by last name †Correspondence: hanliu@northwestern.edu

Paper arXiv 2607.13049 BibTeX Code (coming soon)

SPINE at work

One debugging session on DOBOT X-Trainer. In scenario D-HY2 we unplugged the robot-control Ethernet cable and moved both controller IPs to the wrong subnet.

$ debug start running declared teleoperation pipeline… pipeline FAIL controllers unreachable readiness check FAIL robot-control link blocker ledger: 2 active
Step 1 of 4

Observe

SPINE runs the teleoperation pipeline and the readiness checks. Both fail, and each failure is logged as an open blocker.

live pipelinereadiness checks

Illustrative walk-through, not a recorded log.

Foundation models give robots strong high-level reasoning, but getting a real robot to act still means aligning drivers, networks, sensors, controllers and safety constraints for each platform. We call this integration layer the robot's spinal cord, and it is a main bottleneck for scaling embodied AI. SPINE (Scalable Physical Integration with ageNtic Expertise) is an agentic framework that debugs bimanual robots until teleoperation works. A profile builder compiles what SPINE needs to know about a robot, and a debugger uses that profile to diagnose, repair and verify.

We tested SPINE on two bimanual robots across 12 debugging scenarios. A robotics novice using SPINE resolved more faults, and faster, than human operators using Claude Code. On DOBOT X-Trainer, SPINE resolved every fault (humans 76%) and cut mean time-to-teleoperation by 30%. On AgileX PiPER, it also resolved every fault (humans 99%) and was 38% faster.

SPINE vs. human operators

DOBOT X-Trainer · 7 scenarios

Faults resolved
SPINE100%
Human operators76%
Mean time to teleoperation 30% faster
SPINE10:20
Human operators14:51

AgileX PiPER · 5 scenarios

Faults resolved
SPINE100%
Human operators99%
Mean time to teleoperation 38% faster
SPINE7:48
Human operators12:30

A robotics novice runs SPINE; human operators use Claude Code with the same model (Claude Sonnet 5.0). Times include unsuccessful trials.

DOBOT X-Trainer bimanual teleoperation platform seen from above

DOBOT X-Trainer

Two follower arms, two leader arms, RealSense cameras and a vendor SDK over static Ethernet.

7 SCENARIOS
AgileX PiPER bimanual platform

AgileX PiPER

Four PiPER 6-DOF arms on ROS Noetic with USB-to-CAN adapters and RealSense cameras.

5 SCENARIOS

Method

SPINE runs inside Claude Code as two skills backed by Python. /spine-profile-build learns a robot once, and /spine-debug uses what it learned to get teleoperation working again. The agent chooses each fix, but only the backend can close a case, and only after its checks pass.

Profile building

The operator provides the robot's documentation, a known-good copy of its software and a few setup facts. Nine subagents read these in parallel. Eight each write one part of the profile (identity, components, interfaces, software, configuration, procedures, safety, diagnostics), and the ninth reconstructs the commands that bring up teleoperation. The operator confirms the command sequence, interfaces, hardware playbooks and readiness checks. Nine reviewers and an adjudicator then check the profile, and a deterministic checker validates it before use.

Debugging

debug start runs the teleoperation pipeline and records each failure. Four read-only subagents examine terminal output, software and configuration, hardware visibility and readiness checks in parallel. SPINE applies software fixes itself through debug after-fix. For a physical fix, debug continue walks the operator through one step from a hardware playbook. debug finish closes the case only when the full pipeline and the readiness checks pass in the same run, and a fix enters incident memory only after it passes validation.

SPINE debugging workflow: persistent robot context (robot evidence, robot profile, validated memory) grounds a four-stage loop of observe, triage, repair and verify; unresolved cases return new evidence to triage.
Figure 1. SPINE debugging workflow. Triage runs four read-only subagents in parallel and splits the work into software edits and operator actions. Repair applies one of them. Verification reruns the live pipeline and readiness checks together, and unresolved cases go back to triage with the new evidence.
SPINE system overview: setup inputs feed /spine-profile-build, which produces a robot profile; /spine-debug uses the profile with runtime inputs through start, after-fix, continue and finish, supported by /spine-check and /spine-closeout.
Figure 2. SPINE system overview. The profile builder compiles robot context. The debugger combines it with live evidence and incident memory to fix software, guide physical repairs and check that teleoperation works.

Debugging scenarios

The 12 scenarios implant software-only (SW), hardware-only (HW) or hybrid (HY) faults, two per scenario except D-HW1. An administrator implants them without telling the operator. A scenario is solved when the follower arms and grippers follow the leader arms and every required camera stream is live.

Results

SPINE ran three trials per scenario. The human baseline pools seven operators on DOBOT and nine on PiPER, all using Claude Code with the same model (Claude Sonnet 5.0) but without SPINE. OSS is the fraction of implanted faults resolved in a trial. TTO is time-to-teleoperation, mean ± SD, and includes unsuccessful trials.

DOBOT X-Trainer

SPINE resolved every fault, against 0.76 for human operators, and cut mean time-to-teleoperation from 14:51 to 10:20, 30% faster. Most of the gain came from software-only and hybrid scenarios. Both groups solved all hardware-only ones. SPINE was faster in six of seven scenarios. The exception was D-HY1, where it took 14:55 against 11:55 but resolved every fault, while humans averaged 0.64.

SPINEHuman operators
Time-to-teleoperation

mm:ss, mean with SD, lower is better

Faults resolved (OSS)

mean fraction of implanted faults

Per-scenario table · DOBOT X-Trainer

AgileX PiPER

Human operators already resolved nearly every fault (0.99), so the difference on PiPER is speed. SPINE was faster in every scenario and cut mean time from 12:30 to 7:48, 38% faster. P-HW1 dropped the most, from 13:36 to 2:45.

SPINEHuman operators
Time-to-teleoperation

mm:ss, mean with SD, lower is better

Faults resolved (OSS)

mean fraction of implanted faults

Per-scenario table · AgileX PiPER

Ablation on DOBOT

We removed either the robot profile or the hardware playbook and ran each DOBOT scenario once, so these results are preliminary. Without the profile, SPINE resolved 0.79 of faults and mean time rose from 10:20 to 17:00. It solved half of D‑HW2 and none of D‑HY1. Without the hardware playbook, it resolved 0.57 and took 31:25, about three times as long. Both variants still solved every software-only scenario and failed where the fix needed physical work.

Citation

If you find this work useful, please cite it.

BibTeX
@article{ham2026spine,
  title={SPINE: Bridging the Cyber-Physical Gap with Agentic AI},
  author={Ham, Minkyu and Kim, Dongho and Lee, Chan and Kim, Min Jun and Zhang, Yixi and Wang, Jiayi and Liu, Han},
  journal={arXiv preprint arXiv:2607.13049},
  year={2026}
}

Acknowledgements

We thank Guo Ye, Jihai Zhao, Tony Dong, Jiale Chen, Vincent Cheng, Soyeon Park, Dennis Wu, Hongyu Chen, Donggyun Lee, and Mingxian Wang, who volunteered as human operators in our experiments and/or ran experiments for the SPINE agent. We also thank Guo Ye for the numerous conversations about the design of SPINE, which helped inspire certain elements of it.