Embodied practice 5: communication is back—may transfer resume?

The previous experiment detected stale observations and stopped transfer. The next question is whether communication recovery authorizes an old action to continue.

The revealing result is a counterexample: the receipt-only policy resumed nine times while an old good packet was replayed. Eight resumptions lacked fresh grasp evidence, yet placement eventually succeeded. An endpoint-only score hides the repeated interruptions.

Code and reproduction · Raw evidence and plots · Independent audit

1. Stopping and resuming require different evidence

A communication gap may leave the cube firmly grasped, or conceal a real drop. Transport recovery does not establish the task precondition.

We ask three questions: is this a new observation; does it continuously support a valid grasp; and is the old time-indexed path still appropriate? The first two authorize recovery. The third determines the next command.

2. Frozen setup

The same Cartesian gripper, cube, tray and original motion run in Python 3.12.14, MuJoCo 3.3.7 and NumPy 2.2.6. Commit 8142cdb froze the protocol before the formal run.

Parameter	Value
Physics and control	2 ms
Capture period	20 ms
Fault onset	4.8 s, during transfer
Delivery gap	[4.8, 5.04), 240 ms
Stale threshold	Age reaches 60 ms
Bad-evidence threshold	Span reaches 40 ms
Recovery confirmation	At least 100 ms
Hold deadline	600 ms after stopping
Shared horizon	10.8 s
Matrix	Six conditions × three policies = 18 rollouts

There is one deterministic run per cell, not 18 independent random trials or an estimate of deployment success. Observations are privileged simulator state, not visual perception.

3. Six conditions

Condition	Intervention
clean	No disturbance
delay-40ms	Every capture arrives 40 ms late
reordered	Every third post-onset packet arrives 80 ms late
gap	Suppress delivery during the gap; resume new captures without backlog
stale-replay	After the gap, replay the last good old packet until 5.2 s
gap-and-drop	Open the gripper during the same communication gap

Opening changes actuator commands; contact and gravity move the cube. Its position is not teleported. Delay and reordering do not force an alarm: remaining in motion is itself a result.

4. Three recovery policies

Latched: keep the held target and abort after 600 ms, without resuming.

Receipt-only: any valid delivered packet resumes the original wall-clock path, even when duplicate, stale or inconsistent with a grasp. This is an intentionally weak baseline.

Revalidate: require strictly increasing captures made after the stop; age below 60 ms; bilateral contact; cube height above 0.12 m; cube-to-gripper distance below 0.05 m; and at least 100 ms of evidence with adjacent captures no more than 20 ms apart. A failed condition resets the window.

At the 600 ms deadline, abort takes precedence over confirmation on the same tick. These thresholds belong to this teaching scene, not a universal robot safety specification.

5. Why replay causes repeated stop–resume cycles

The following times come from the stale-replay logs:

Time	Event
4.782 s	Last pre-gap good capture
4.842 s	Its age reaches 60 ms and transfer stops
5.042 s	Receipt-only resumes on that same capture, now 260 ms old
Afterwards	Staleness stops motion again; another delivery reopens it
5.302 s	Revalidation first resumes after 100 ms of fresh evidence

Receipt-only resumes nine times, eight on stale evidence. Inspect recovery events together with the final task outcome.

This plot comes from control logs. It is neither a camera image nor a physical robot recording.

6. Why replan the remaining path?

Physics and the clock keep advancing while a target is held. Looking up the old path at the current wall-clock time can skip an unexecuted segment.

Revalidation creates a 1.5 s transfer from the current measured gripper position to the original transfer destination, followed by the original lowering, release, retreat and settling stages. Holding a target does not freeze physical state.

The policy changes both authorization and trajectory generation. This comparison therefore evaluates a bundle. A separate ablation would be needed to attribute differences to the 100 ms gate alone.

7. All 18 outcomes

Condition	Latched	Receipt-only	Revalidate
Clean	Placement succeeds	Succeeds	Succeeds
Fixed 40 ms delay	Succeeds	Succeeds	Succeeds
Every third packet 80 ms late	Succeeds	Succeeds	Succeeds
240 ms gap	Aborts, no placement	One resume, succeeds	One resume, succeeds
Stale good replay	Aborts, no placement	Nine resumes, eight unsupported; succeeds	One resume, succeeds
Gap and forced opening	Aborts, no placement	Six unsupported resumes; fails placement	No resume, aborts

“Unsupported” means capture age at least 60 ms or a false grasp predicate at resumption. It is not a claim that a physical robot would necessarily become unsafe.

Revalidation resumes at 5.142 s for the plain gap and 5.302 s for stale replay. After forced opening, it aborts at 5.442 s. Failing placement in the latter case is compatible with correctly refusing an unjustified continuation.

8. Checks independent of the decision class

The audit reads saved CSV/state files without calling Gate:

• It recomputes endpoint outcomes for 18 rollouts.
• Twelve policy prefixes before their first decision divergence match exactly: maximum absolute qpos/qvel/ctrl error is zero.
• Both revalidation windows satisfy ordering, age, duration and grasp predicates.
• It records SHA-256 for 108 raw files and checks packet accounting and deadlines.
• Eighteen new boundary tests and 41 existing tests pass.

This checks evidence consistency and logic boundaries; it is not validation against another physics engine.

9. Reproduce and inspect

From the embodied repository, use a new output directory:

.venv/Scripts/python experiments/vl01_recovery_gate/test_gate.py
.venv/Scripts/python experiments/vl01_recovery_gate/run.py --out evidence/my-recovery-run
.venv/Scripts/python experiments/vl01_recovery_gate/audit.py evidence/my-recovery-run
.venv/Scripts/python experiments/vl01_recovery_gate/render.py evidence/my-recovery-run

Each control.csv records decisions every 2 ms. State and trajectory samples are saved every 20 ms. Events record stops/resumes/aborts; replans record new starting states.

For historical evidence checked out on another platform, run portable-audit.py evidence/recovery-20260930. It restores only an LF/CRLF representation matching the recorded hash in a temporary directory, then runs the frozen audit. Original evidence remains unchanged.

10. Limits and the next question

Within these fixed conditions, delivery alone is insufficient authorization. Sustained new grasp evidence plus remaining-path planning separates communication interruption from lost grasp.

Real networks, clock skew, vision error, different objects and different fault phases remain untested. This is not regrasping, a ROS2 bridge or completion of M4/M5. The next useful experiment is an ablation separating the gate from replanning.

For simulator stepping and state basics, see the MuJoCo 3.3.7 simulation documentation. The code, frozen protocol and raw logs define the scope of this article.