ego-handpose

# EgoDex/EgoVerse EEF14 Action Cache This repository contains Ego2Robot-style 14D end-effector action targets for the EgoDex and EgoVerse samples in `ego-200h-release-v1`. It intentionally does not duplicate RGB videos or Wan VAE latents. The referenced VAE cache is: ```text ego-200h-release-v1/vae_cache dataset_signature: f95c52586266dbc6eb7cca152376b9a48111e538a76561f8b366e1a0c244f219 ``` ## Alignment `eef_action_cache/logical_identity.pt` contains `vae_indices`. For EEF sample `j`, load the original VAE sample at `vae_indices[j]` and the EEF target at sample `j` in this action cache. The action-cache shards contain 4,096 samples each. ```python import json from pathlib import Path import torch eef_root = Path("eef_action_cache") manifest = json.loads((eef_root / "manifest.json").read_text()) identity = torch.load(eef_root / "logical_identity.pt", weights_only=True) j = 0 vae_index = int(identity["vae_indices"][j]) shard_size = int(manifest["shard_size"]) payload = torch.load( eef_root / f"shard_{j // shard_size:06d}.pt", weights_only=True, ) action = payload["action"][j % shard_size] # [32, 14], float32 action_mask = payload["action_mask"][j % shard_size] # [32, 14], bool ``` Each 14D row is: ```text left [delta xyz (m), delta rotation-vector (rad), delta gripper width (m)] right [delta xyz (m), delta rotation-vector (rad), delta gripper width (m)] ``` Actions cover frames `anchor..anchor+31`. The paired VAE sample covers video endpoints from `anchor..anchor+48`. Invalid or unavailable hand dimensions are marked by `action_mask`; ego samples contain no robot proprioception. The cache contains 748,123 samples: 330,158 EgoDex and 417,965 EgoVerse. Normalization statistics are provided under `normalization/`.