Before using SARM (Stage-Aware Reward Modeling) to weight demonstrations, inspect its selected prediction head and sampling stride against the episode frames. An interpolated trace is not a fresh model judgment on every frame, and a learned progress score is not independent confirmation of task success.
The current LeRobot SARM guide, checked October 15, 2026, separates prediction visualization from optional Reward-Aligned Behavior Cloning, or RA-BC. The latter weights training samples from predicted progress improvement. This is a documentation review of the source-install main workflow; no reward model, visualization or weighted policy was run.
Keep annotation mode beside the plotted head
The guide distinguishes single_stage, dense_only and dual annotation modes. In single_stage, the whole episode is one stage with a linear time-based progress target. The other modes use dense annotations, or both sparse and dense annotations, to define different granularities. Record that target construction before interpreting a plotted curve as evidence of successful behavior.
Choose the prediction head explicitly when comparing traces. A sparse-stage estimate and a dense-stage estimate can answer different questions even when both are displayed on a normalized scale. Keep the annotation revision and any annotation-derived target alongside the prediction; neither should be silently renamed as measured physical completion.
Locate the frames that were actually evaluated
The documented stride computes predictions every N frames and interpolates the rest. The visualization can show annotation-derived targets when available with stride 1. Its visualize-only option inspects predictions without writing a progress parquet file.
Consider an illustrative episode in which an object is briefly released and regrasped between sampled frames. A smooth interpolated curve may conceal that event. This is a possible diagnostic failure, not a reported SARM result. Mark computed versus interpolated points, then inspect the surrounding frames and compare a denser prediction pass for that interval.
For a proposed acceptance record, retain the dataset revision, episode/frame identities, model revision, annotation mode, head and stride. Include examples with hesitation, corrective motion and failure, rather than checking only smoothly successful demonstrations.
Trace a weight back to its two progress values
RA-BC uses a predicted progress difference across the policy’s chunk_size interval. Inspect both endpoint predictions when a segment receives an unexpected weight. Changing the head, stride or chunk size can change the quantity being compared even when the underlying demonstration remains the same.
Before using those weights for training, trace representative high, low and zero-weight segments back to their frame identities and inspect them against the video. This is proposed validation, not an executed weighting audit. The reward-model guide explains why learned rewards remain proxies; this narrower check makes the SARM sampling and weighting path inspectable before it changes policy training.