Predict several actions at once.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Start with the task

Planning three pushes at once can reduce decision calls. A bump after the first push can also make the remaining two commands stale.

Mechanism and units

Six network outputs encode commands at offsets zero, one and two. Targets are consecutive teacher commands, with the last available command repeated at the episode boundary. A horizon of three executes all three outputs; a horizon of one executes only the first and requests another observation. Each command lasts 0.1 seconds.

A three-command block spans 0.3 seconds. If a disturbance occurs just after its first command, two later commands can be issued without seeing the changed position. Fewer calls therefore trades feedback frequency against execution work. The model can be frozen throughout; observing again is not retraining.

Work through the evidence

Inspect the selected receipt immediately after the ninth interval. Compare observedAt between single-step and chunk modes. Count observations and learned-policy calls, then compare full-task success and final error. Keep drag, delay and random offsets matched.

Compare horizons one and three using the same trained weights and displacement.

What this experiment can establish

The interface reports simulated seconds, not real latency or energy. Chunk predictions are learned, but the scheduler choosing how many to execute is programmed. These public comparisons do not establish real-robot timing safety.

Keep collection, fitting, frozen execution and interpretation as separate records. Save the configuration and complete result, including failures, in the robot-learning project. Report submission checks required evidence and text presence; reasoning and implementation quality need independent review.