HadiZayer/droid-normals

DROID camera-space normal control videos

Robot surface-normal videos for the external cameras of selected DROID episodes (from the task-selection list: 14 official task categories + 6 lab-defined tasks). Every video frame is rendered (no temporal downsampling), 1280x720.

Layout

episodes/<uuid>__<cam_serial>/
  reference.png          # first frame of the RGB video
  rgb.mp4                # original DROID external-camera video
  control_normals.mp4    # normal map video (H.264, yuv444p, CRF 8; LOSSY)
  mask.mkv               # binary mask video (FFV1, lossless; 255 = robot rendered)
  meta.json              # task, outcome, losses, intrinsics, world2cam, frame mapping
manifest.csv             # one row per camera

One folder per camera, not per episode: an episode contributes one folder for each external camera that passed the filter below.

How it was made

• Robot: Franka Panda + Robotiq 2F-85 URDF, driven by trajectory.h5 joint_positions; finger_joint = gripper_position * 0.725.
• Camera: refined_extrinsics (world2cam) and measured intrinsics (fallback VGGT) from the refined camera JSONs, intrinsics scaled from 640x360 to the video size.
• Filter: a camera is included only if it has refined_extrinsics and its per-camera optimization loss is <= 0.1. Cameras failing this are skipped individually.
• Frame mapping: traj_idx = min(int(video_idx * traj_len / n_video_frames), traj_len - 1).

Normal encoding

• OpenCV camera space, rotation only: +X right, +Y down, +Z into the scene, so a camera-facing surface has n_z < 0.
• uint8 = round((n + 1) / 2 * 255) per channel. Background is (0, 0, 0), which would decode to (-1, -1, -1), so always gate on mask.mkv.
• control_normals.mp4 is lossy (small error inside, larger at silhouette edges); decode, then renormalize to unit length:

import cv2, numpy as np
ok, bgr = cv2.VideoCapture("control_normals.mp4").read()
_, m = cv2.VideoCapture("mask.mkv").read()
n = bgr[..., ::-1].astype(np.float32) / 255 * 2 - 1
n /= np.maximum(np.linalg.norm(n, axis=2, keepdims=True), 1e-6)
n[m[..., 0] <= 127] = 0

rgb.mp4 is the unmodified original video from the DROID dataset (https://droid-dataset.github.io, CC BY 4.0; Khazatsky et al., 2024). Normals/masks are derived from DROID trajectories and camera calibration.