DEV Community

liao-zhizhi
liao-zhizhi

Posted on

From Code to Motor: Why Software Audit Sidecars Fail in Embodied AI

#ai #robotics #systems #opensource

In the digital realm, we have tools to catch "false completions" from LLM agents. We use software sidecars to record independent physical witnesses (like file hashes or git diffs) and output four-valued verdicts: PASS, FAIL, INVALID, NOT_RUN.

But what happens when the agent is not writing code, but controlling a drone or a robot?

Recently, I formalized the structural faults that occur when we move digital audits into embodied systems. The core issue is simple: A software exception can roll back a file write, but it cannot roll back a crash.

When a software audit sidecar is applied to physical agents, it faces three structural faults:

  1. Fault I: Shared-scheduler refusal
    If the sidecar shares the same CPU and fault domain as the flight controller, it dies when the controller deadlocks. Refusal remains an exception, which is unreachable on a physical timescale. Hardware isolation (like an independent hardware watchdog or MCU) is required to enforce the refusal.

  2. Fault II: Physical-oracle substitution
    The physical world is nearly a free outcome oracle (e.g., wind resistance, inertia). However, software sidecars often substitute this with a homologous shadow—reading logs, self-reports, and software hashes. When the auditor and the audited share the same clock or sensor channel, the temporal witness is colluded away.

  3. Fault III: Assurance-coverage collapse
    A Lean 4 development can perfectly close the discrete four-valued lattice. But physical systems require proving worst-case execution time (WCET), allocation freedom, and sensor timing. The fraction of discharged physical objectives collapses toward zero. Formal logical completeness does not equal real-time physical assurance.

The Remaining Architecture: Split the Layer
The runtime safety of embodied AI cannot just be "another Python sidecar." It requires a split between:

A slow record layer (can stay in Python).

A fast enable layer (must run on an independent MCU in C/Rust). The default on silence or packet loss must be a predetermined safe state (Fail-Safe).
Enter fullscreen mode Exit fullscreen mode

I just published the full paper on this topic: The Physical Grounding Fault: Why Software Audit Sidecars Fail in Embodied Agents.

This is Paper 7 in a continuous research chain that moves from formal verification (EquivSVA) to information-theoretic limits of self-evaluation, and finally to physical grounding.

You can read the paper (DOI: 10.5281/zenodo.24139787) and explore the full open-source progression (including the sidecar tools) here:
👉 GitHub Repo: https://github.com/liao-zhizhi/causal-еvolution-paper
👉 Paper Archive: https://doi.org/10.5281/zenodo.24139787

(Disclaimer: This research is restricted to civil special-use robotics and unmanned aircraft safety. It does not involve weaponization or tactical systems.)

Top comments (0)