A navigation policy can receive a new action while still using an old interpretation of the scene. To evaluate that mismatch, measure how old the semantic input is when an action is issued, keep current robot observations available, and report outcomes across delay ranges. Average inference time alone cannot show whether a robot acts sensibly after its reasoning falls behind.
Consider a robot approaching a corridor intersection. Its camera sees someone step into the path, but the language model's description still refers to an empty corridor. This is an illustrative failure scenario: the question is whether the action policy can use the new observation while recognizing that its semantic advice is stale.
Timestamp the scene that produced the advice
The TIC-VLA research project describes a slow reasoning VLM and a fast action expert. The action expert receives current observations and robot state alongside delayed semantic features and latency metadata. Its training injects reasoning delays during imitation learning and reinforcement learning. These are the authors' methods, rather than tests performed for this article.
For a local evaluation, distinguish three times: when the source observation was captured, when semantic inference finished, and when the resulting action was issued. Semantic age at action time is the last timestamp minus the first. Completion time minus capture time measures only the reasoning portion; queuing and later reuse can make the advice older still.
For an invented common-clock trace, result R1 uses an observation captured at 0 ms and becomes ready at 500 ms. The action layer uses R1 twice:
| Action issued | Semantic result used | Current camera observation |
|---|---|---|
| 800 ms | R1 from 0 ms: semantic age 800 ms | Frame from 750 ms: camera age 50 ms |
| 1,600 ms | The same R1: semantic age 1,600 ms | Frame from 1,550 ms: camera age 50 ms |
The reasoning completion interval was 500 ms in this example, yet the second action uses advice that is 1,600 ms old. Fresh camera input does not make the reused semantic result new. These times and identifiers are illustrative logging choices, not measured TIC-VLA output or guaranteed release fields, and neither age is proposed as a deployment limit.
A log containing only the model's response time loses this distinction. Record a semantic-result identifier with each action so that repeated use of one result remains visible. Choose a common clock or document how timestamps from different machines are reconciled.
Delay training inputs without hiding fresh observations
The evaluation should ask whether training resembles the asynchronous system the robot will actually use. An artificially delayed semantic feature is a different intervention from delaying the camera, proprioception, and semantic feature together. The latter also removes the fast loop's chance to react.
A proposed comparison would hold the checkpoint family, scene set, action interface, and observation availability fixed, then compare matched-delay training with a baseline trained under temporal alignment. Include variation in delay, rather than assuming every inference takes exactly the mean measured time. This isolates a training question; it does not establish a safe maximum delay for every robot.
Read the failures within each delay range
Keep task completion alongside elapsed time, stopped episodes, and the chosen collision or clearance measures. Report the number of episodes in every delay range. A long-delay bin containing only easy routes can make compensation look stronger than it is.
- Separate reasoning age from camera age and network latency.
- Compare the same route and crowd conditions across policies where the protocol permits.
- Preserve episodes that terminate early; explain their outcome labels.
- Inspect whether repeated old advice precedes a wrong turn or a failed reaction.
The project's DynaNav evaluation uses dynamic navigation scenes, while its real-robot demonstrations provide a separate evidence layer. Neither supplies a universal deployment envelope. For the broader action-model distinction, see what a VLA model controls.
Before accepting a latency claim, request the semantic-age distribution and a few synchronized observation/advice/action traces. Those artifacts reveal whether the policy coped with old reasoning or merely completed runs where reasoning happened to arrive quickly. Research source checked October 14, 2026.