A faster action loop cannot recover motion that the policy never observed. If a robot reaches behind a moving object even when its camera frame is fresh, test whether its input describes the object’s direction and speed. If it reverses or hesitates at a familiar-looking stage of a task, test whether it knows what it has already commanded. Those are different missing pieces of temporal context, and they call for different probes.
Picture a bottle handover. At one instant, the bottle occupies the same pixel position whether the person carrying it walks left or right. A single image can locate it but cannot distinguish those two futures. Shortening inference time gives the robot a newer location; it does not derive direction from that image. The TEMPO paper studies this distinction. The engineering question is: Which information is absent when this policy fails?
A late command and a missing motion cue leave different traces
Preserve the camera timestamp, inference start and end times, action dispatch time, and a synchronized target-position trace. A rollout video shows a miss; the trace helps connect it to elapsed time or target motion. The patterns below are diagnostic hypotheses, not TEMPO results.
| What you observe | First comparison to make | More useful next probe |
|---|---|---|
| The gripper aims at the target’s old location, and misses grow as the observation-to-action delay grows | Plot miss direction and distance against measured delay while holding the target path as similar as possible | Reduce delay or compensate for it, then repeat the same trajectory |
| The gripper chooses the wrong side even when delay is low; opposite target directions produce similar commands from similar frames | Pair moments with a similar current target position but opposite recent motion | Add recent visual positions or a motion feature while keeping the timing path fixed |
| The gripper stalls or alternates between reaching and withdrawing near a task transition | Compare nearly identical frames from before and after grasp, release, or return | Supply a logged task-phase cue or recent robot-command history and check whether the two moments separate |
These patterns can coexist. Old frames and missing motion history may both matter. One clip cannot assign a single cause. For broader evaluation controls, see the robot VLA evaluation guide.
Keep timing fixed while testing what the policy knows
For a handover test, have the bottle cross the same position from left and right at several speeds. Fix the robot start, camera, instruction, and action limits. Log the frame consumed and command received. A history-enabled policy can be compared with a single-frame policy only if control rate, training examples, and evaluation starts remain matched.
Separate tracking from commitment. Before contact, does the gripper follow the bottle’s travel or its latest detected position? At the grasp, does it commit or repeatedly reverse? These failure labels point toward motion cues, phase cues, or execution timing.
This is a proposed test, not a result from our lab. History length and acceptable delay require measurement on the intended system. If failures follow inference spikes in an almost static scene, address execution timing first.
What the matched TEMPO comparison actually shows
TEMPO adds two inputs to a pretrained VLA: a representation of recent scene motion from video features, and a compact record of the robot’s recent commands. In the paper’s Table 4, the VLASH baseline completed 38% of 50 real-world handover trials; TEMPO completed 74%. The 44% figure in the same table belongs to a different baseline, RTC, and should not be used as the starting point for the VLASH-to-TEMPO comparison.
On this apparatus and task, temporal inputs helped a policy that already used asynchronous inference. The result does not establish transfer to another robot.
Recent commands can identify a phase, but they are not measured state
The second input matters when two visually similar moments require opposite actions. A gripper near a bottle may be approaching to grasp, or returning after release. If the policy sees only the current image, both can look plausible. A compact history of its own commands can help distinguish which phase it has entered.
A history of commands is not proof that the robot executed them. If a joint stalls, a grasp slips, or a low-level controller clips the command, the command trace can describe the intended phase while the physical state is elsewhere. A local test should therefore retain measured joint and gripper state, task events, and the command stream as separate records. Whether to feed measured state into a different design is a new experiment, not a claim established by TEMPO.
The adjacent question of whether a perception or reasoning output contains enough information to specify an action is covered in the GIFT action-sufficiency analysis. Here the decision is more specific: does the same scene require a different action because of what just moved or what the robot just did?
A useful stopping point for the diagnosis
After a bounded set of matched runs, write down three distributions: observation-to-command age, target-direction error at the intercept, and the count of phase reversals near grasp or release. Save representative successes and failures with the same timestamps. If delay explains the errors, work on execution timing. If opposite motions remain indistinguishable at matched delay, test visual history. If action reversals cluster at repeated-looking stages, test a phase cue and verify it against measured robot state.
Those records identify which input or timing change to test next. No robot experiment was run for this article. Source checked September 29, 2026.