Research

Action Chunking under Low Data

Sep 2026 – Present

An ongoing study comparing observation history and prediction memory under limited demonstrations and hidden command delay, with an audited evaluation pipeline.

Schematic placeholder cover for the action-chunking study

Research Question

Under limited demonstration data and hidden command-application delay, does a policy benefit more from observation history or controller-level memory of overlapping action predictions? The working title is “Observation Memory or Prediction Memory under Hidden Delay?” The mechanism study is ongoing.

Why It Matters

A robot may observe the present while earlier commands are still waiting to execute. Observation history and aggregation of past predictions offer different ways to handle that mismatch. A controlled comparison must separate them from prediction horizon, the number of actions executed before replanning, and data volume.

My Role and Research Process

I am developing the protocol and training/evaluation infrastructure, organizing paired scenarios and nested demonstration subsets, and checking controller behavior and reproducibility. The process starts with a reference baseline and auditable controls before testing the memory hypotheses.

Study Design

The main protocol uses image-based Diffusion Policy and Drake Push-T. It fixes executed prefix E = 1 and prediction horizon H = 16, varying observation history C, aggregation window W, demonstration count N, and hidden FIFO command delay D. Nested subsets use N = 25, 50, 100, and 180. Generation evaluations (NFE), executed prefix E, and command delay D are distinct variables.

Paired scenario manifests and sampling seeds support matched comparisons. DBP and controlled ACT-CVAE, ACT-no-VAE, and Drifting-ACT comparisons are proposed or gated extensions, not completed findings. The memory hypotheses remain hypotheses.

Findings and Evidence

Implemented work includes upstream training/evaluation integration, paired scenario manifests, nested subsets, history and aggregation controllers, FIFO delay tests, and audit logging. The Stage0 record reports 47 local tests and two byte-identical reproducibility runs of two rollouts each. These are engineering checks, not evidence of a memory mechanism’s advantage.

The provisional upstream-reference evaluation compared final and best EMA checkpoints on the same 50 scenarios and sampling seeds: 41/50 (82%) versus 38/50 (76%), from one training seed. This baseline used C2/rawH16 with 15 future actions and E8; it is not the main H16/E1 experiment. It does not establish a memory-type improvement, a multi-seed conclusion, or a new method outperforming the baseline.

Limitations and Reflection

The full mechanism study is unfinished. Reference-checkpoint results come from one training seed; small reproducibility checks do not establish robust task performance. Historical incomplete pilots are not treated as completed cross-seed results. Controller semantics and evaluation provenance must be fixed before interpreting performance differences.

What I Can Do Next

Complete controlled H16/E1 mechanism experiments, repeat across training seeds, and examine paired outcomes across data sizes and delays. Architecture extensions should follow their protocol gates. These are future experiments, not measured improvements.

Source note: repository experiment design v0.3, September 23 Stage0 and Drake smoke records, and the paired baseline summary. Source repository (private; collaborator access required). Verification is reported from those records; this page does not claim to have rerun the research experiments.