sinew · plots · r8

R8 — v2f_109 force tracking

Visual evaluation of the v2f_109 force predictor against actual_clean wrenches on the fmb_real/val split (shape 3, K=1 single-frame).
Production checkpoint upgraded to v2f_122 on 2026-05-24. These plots reflect the R8 baseline (v2f_109). Newer evaluations against v2f_122 supersede this for production decisions. Kept here for historical reference and to anchor the F1 / MAE regression test.
Reading the plots. Time-series traces compare predicted (orange) vs ground-truth actual_clean (blue) wrench components Fx, Fy, Fz over the episode timeline. Per-shape MAE bar aggregates across the val split; the scatter shows direction-cosine accuracy as a function of force magnitude.

Summary

checkpoint
v2f_109_3shape_optY_seed7
val split
fmb_real/val
temporal window
K=1 (single-frame)
shape
3 · 4 episodes shown

Per-shape MAE (shape 3, mean over val)

component MAE (N)
Fx2.710
Fy2.932
Fz3.390

Aggregate plots

Per-shape MAE bar — F_x, F_y, F_z, shape 3
Per-shape MAE bar. Mean absolute error per force component, shape 3, val split.
Scatter: direction-cosine accuracy vs force magnitude
Direction-cosine vs ‖F‖. Predicted-to-ground-truth direction agreement plotted against ground-truth force magnitude. Higher cosine = better direction recovery; trend at low ‖F‖ reveals where the predictor loses directional signal.

Episode time-series (shape 3)

Episode 0 time-series — predicted vs actual_clean force
Episode 0. Predicted vs actual_clean for Fx, Fy, Fz.
Episode 1 time-series — predicted vs actual_clean force
Episode 1. Predicted vs actual_clean for Fx, Fy, Fz.
Episode 2 time-series — predicted vs actual_clean force
Episode 2. Predicted vs actual_clean for Fx, Fy, Fz.
Episode 3 time-series — predicted vs actual_clean force
Episode 3. Predicted vs actual_clean for Fx, Fy, Fz.

Source data

artifactpath
index manifestindex.json (this directory)
checkpointoutputs/v2f_109_3shape_optY_seed7/checkpoints/predictor_final.pt
val data/home/dom_iva/fmb_real/val (DL_A6000)
NAS FTP archiveftp://143.248.121.169:7002/IntelligentManipulationTeam/DomrachevIvan/ · path under r8_force_comparison/