← Paper Digest

Camera equalisation of an 11.5x drop in environmental illuminance: four imitation-learning policies on an SO-101 arm

Simon Cowling

November 2026

TL;DR Four imitation-learning policies (ACT, SmolVLA, Diffusion Policy, pi0.5) on an SO-101 arm were run at about 520 and 45 lux, and success went from 20/40 to 19/40. Automatic exposure left the front camera's images only 3% darker, so lux alone does not describe what a policy sees.

What they do

A real-robot check of a common assumption in policy benchmarks: that logging room illuminance in lux describes the lighting a policy sees. Four policies trained on the same 50 teleoperated demonstrations are run at the recording light level and at about one eleventh of it. Every recorded frame is then measured to see how much darker the images actually got.

How it works

  • Rig and task. A 6-DoF SO-101 arm on LeRobot with a fixed front camera and a wrist camera (640x480, 30 fps). The task is to put the green ball in a basket while ignoring two red balls, within 30 s.
  • Policies. ACT, Diffusion Policy, and SmolVLA and pi0.5 fine-tuned from their pretrained bases, all with LeRobot reference configurations.
  • Protocol. One block of ten fixed ball placements per policy and light level, on consecutive days: 519 lux average at the workspace versus 45 lux (artificial lights off, daylight screened). Brightness, contrast, sharpness, saturation and white balance of every frame are compared with 20 training episodes.

Results

  • Success was 20/40 in normal light and 19/40 in low light: -2.5 points (95% CI -23 to +19), with no evidence the policies differed (interaction p = 0.86). No policy changed by more than one success in ten. A policy-averaged drop of 20 points or more has posterior probability 0.024; a 10-point drop cannot be excluded.
  • Lux fell about 10x at both camera positions, but mean image brightness fell only 3% (front) and 19% (wrist). Both cameras were on automatic exposure and white balance.
  • Normal-light images, taken at the recording lux, were further from the training images in colour (47% more saturated) than the low-light ones.
  • An accidental repeat block reproduced 7/10 but matched the first block at only 6 of 10 placements, consistent with chance.

Limitations

Ten rollouts per cell detect only very large per-policy changes. Light level is confounded with day and run order, scoring was done live by one unblinded operator, and there was no locked-exposure control, so the role of automatic exposure is inferred. One rig, one task, one checkpoint per policy, and the policies were never actually shown dark images.

Figures

From Figure 2: First frames of a training episode and of placement N00 in the normal-light and low-light ACT blocks; ...
From Figure 2: First frames of a training episode and of placement N00 in the normal-light and low-light ACT blocks; front camera left, wrist camera right. The low-light frames are barely darker than the others.
From Figure 1: Success by policy and light level with 95% intervals; 80 rollouts, 10 per cell.
From Figure 1: Success by policy and light level with 95% intervals; 80 rollouts, 10 per cell.
Abstract

Real-robot benchmarks of learned visuomotor policies often record room illuminance as the lighting condition. We show that, on a typical low-cost rig, room light and image light can be almost unrelated. Four imitation-learning policies (ACT, SmolVLA, Diffusion Policy and pi0.5) were trained on the same 50 demonstrations of a pick-and-place task on an SO-101 arm. Each was run for ten rollouts at about 520 lux and again at about 45 lux, an 11.5x reduction. Success was 20/40 in normal light and 19/40 in low light: a change of -2.5 percentage points (95% CI -23 to +19), with no evidence that the policies differed in sensitivity (interaction p = 0.86). The recorded video shows why: mean image brightness was only 3% lower in low light on the front camera and 19% lower on the wrist camera. Both cameras were set to automatic exposure and white balance, which most likely absorbed the difference; exposure was not locked in this experiment. An accidental repeat of one 10-rollout block reproduced its success count (7/10) but matched the first block at only 6 of 10 placements, consistent with chance. We conclude that lux alone does not describe what a policy sees. Camera settings, and image statistics compared with the training recordings, are the quantities to report and control. Data and code are public.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.