MuJoCo observation loss: zero force is not a stop

git clone https://github.com/VirtualSelect/hohoo-embodied-agent.git hohoo-embodied-agent-study
cd hohoo-embodied-agent-study
git checkout 1d1faab57e20b352761e9abc736c6d277b505e6b
python -m venv .venv
$py = '.\.venv\Scripts\python.exe'
& $py -m pip install -r requirements-lock.txt
& $py experiments/planar_reliability/run.py --out outputs/my-run
& $py experiments/planar_reliability/audit.py outputs/my-run

Pinned code · Archived evidence · Study protocol

Communication stops while the controller continues using its last position and velocity estimate. Does commanding zero force constitute a safe stop? In this frictionless planar model, zero force means no further acceleration, not zero velocity.

Using the development-selected parameters, this experiment compares three stale-observation policies. MuJoCo advances every physical state; nothing teleports the actuator or clears its velocity to manufacture a stop.

Which clock detects staleness?

Each sample carries capture time and becomes usable only after arrival. Age is decision time minus capture time. Reusing the same sample cannot refresh its age.

sample -> 40 ms delay queue -> arrived observation -> estimate/control
                                    |
                           now minus capture time

Delay is 40 ms, position noise 3 mm, control period 20 ms and physics step 2 ms. Policies activate strictly above 100 ms; the exact boundary remains normal control. Conditions are normal delivery, dropped captures from 0.4 to 1.0 seconds, and permanent loss after 0.4 seconds. Three seeds, conditions and policies give 27 episodes.

Noise is drawn even for dropped samples, then discarded. This preserves matching random-number consumption across policies and dropout conditions.

Three policies change different commands

Policy	Age at most 100 ms	Older observation
hold	PD and force clipping	Continue PD from last observation
zero	PD and force clipping	Command zero force
fade	PD and force clipping	Linearly reduce clipped force, reaching zero at 200 ms

Hold does not hold the current physical position: it keeps a stale control computation. The actuator moves while the observed error remains fixed, potentially driving sustained acceleration.

Zero and fade are not closed-loop braking. They add no fresh speed information, friction or brake model. The implementation never clears qvel to obtain a stationary picture.

Temporary loss and permanent loss differ

The following values average final-300ms position RMS over three seeds, in millimetres:

Communication	hold	zero	fade
Normal	2.04	2.04	2.04
Burst loss, 0.4–1.0 s	2.02	2.06	2.05
Permanent loss after 0.4 s	16,025.96	2,748.98	2,370.92

Normal results match because age never triggers the policies. After temporary loss, new observations allow low terminal position error again; terminal averages do not describe the intermediate path.

With permanent loss, hold keeps pushing, zero permits momentum-driven coasting, and fade changes force withdrawal without guaranteeing braking. None of the 27 episodes passes the full position-and-speed contract, including normal episodes with small position error. There is no safety winner in this table.

Why retain such a simple counterexample?

Sending a zero command, stopping controller software, entering actuator braking mode and confirming a stationary object are different events. Verifying only the first cannot justify the fourth. Hardware systems also require actuator semantics, watchdogs, gravity/contact handling, brakes and appropriate safety validation; none is present here.

The 100/200 ms thresholds are teaching parameters, not safety standards. Mass, initial velocity, target and failure phase alter displacement. Testing one loss onset cannot establish a worst-case stopping distance.

Read the trace

For permanent loss under zero, locate the first zero fx/fy row and inspect subsequent vx/vy. Velocity persists in this undamped model; the independent F=m*a check verifies it. An immediate stop suggests state clearing or additional damping in the scene.

For burst loss under fade, inspect capture times after recovery: receiving an old sample again is not fresh sensing. A future braking design must specify available observations and actuator capabilities rather than secretly reading simulator truth to replace missing sensors.

Reproduce the protocol

Run from a fresh checkout in PowerShell. The output directory must not exist. All three articles in this study share one frozen protocol; run it once and inspect the section relevant to each article.

git clone https://github.com/VirtualSelect/hohoo-embodied-agent.git hohoo-embodied-agent-study
cd hohoo-embodied-agent-study
git checkout 1d1faab57e20b352761e9abc736c6d277b505e6b
python -m venv .venv
$py = '.\.venv\Scripts\python.exe'
& $py -m pip install -r requirements-lock.txt
& $py -m unittest discover -s experiments/planar_reliability -p "test_*.py"
& $py experiments/planar_reliability/run.py --out outputs/reliability-my-run
& $py experiments/planar_reliability/audit.py outputs/reliability-my-run

Tested on Windows with Python 3.12.14 + MuJoCo 3.3.7 + NumPy 2.2.6; 3 unit tests, 153 episodes and 306000 rows. The independent audit should print PASS. Other operating systems have not been rechecked. To inspect the archived result without rerunning the experiment, pass evidence/planar-reliability-20261011 to the same audit command. Source hashes normalize LF; evidence hashes use original bytes. Do not rewrite the manifest to conceal a mismatch.

If an import fails, verify that installation and execution use the same $py. Preserve failed outputs and choose a new directory for reruns. Numeric comparisons use the frozen tolerances; timing is a local measurement, not a required speed threshold.

Code and writing are AI-assisted; the findings are limited to the archived, auditable runs. No new paid model requests were made, and this does not claim author-performed manual review.