Hannibal52Barca/demo-end-effector-space

demo_action_space

Derived from adityx23/icl-demo-dataset (lerobot v2.1 format, 285 episodes / 254,171 frames / 27 tasks). Every existing column, task, episode flag (success/valid/keep), and episode_uid is carried through unchanged.

Sibling dataset: demo_joint_space adds the same episodes' joint-space IK targets instead of Cartesian poses. Same source, same episode indices, same pipeline.

What's added

Two new features, observation.left_ee / observation.right_ee (float32, shape [9], names r00, r01, r02, r10, r11, r12, x, y, z): the Cartesian end-effector pose of each arm, forward-kinematics'd from that frame's recorded observation.state (real joint encoders) through the same reference kinematic model demo_joint_space's action.q_target is solved against. This is the FK counterpart of that IK reconstruction — no iterative solve, no warm-starting: observation.state is already a fully-determined joint configuration, so this is a single closed-form forward-kinematics pass per frame.

Rotation representation: 6D, not quaternion. Orientation is stored as the first two rows of the 3x3 rotation matrix, flattened (r00..r12), per the continuous 6D representation of Zhou et al., "On the Continuity of Rotation Representations in Neural Networks" (CVPR 2019) — chosen over a quaternion because it has no antipodal (q / -q) ambiguity and no gimbal-lock-style discontinuity, which matters for anything training a regression head on this pose. Recover the full rotation matrix if needed:

import numpy as np

def rot6d_to_matrix(r6):  # r6: (..., 6) -> (..., 3, 3)
    r0, r1 = r6[..., 0:3], r6[..., 3:6]
    b0 = r0 / np.linalg.norm(r0, axis=-1, keepdims=True)
    r1 = r1 - (b0 * r1).sum(-1, keepdims=True) * b0
    b1 = r1 / np.linalg.norm(r1, axis=-1, keepdims=True)
    b2 = np.cross(b0, b1)
    return np.stack([b0, b1, b2], axis=-2)  # rows

The Gram-Schmidt step is defensive (this pipeline's own rows are already unit-norm and orthogonal to float precision, verified against the quaternion path bit-for-bit) — safe to skip if you trust the source, keep it if consuming rows produced by anything else.

What's NOT corrected

Unlike action.q_target in demo_joint_space, the poses stored here are raw FK of the real encoders — they are not corrected against the recorded action targets. Comparing observation.{left,right}_ee to action.{left,right}_ee directly will show the same per-episode/per-arm rigid offset documented there (near-zero in X/Y, tens of cm in Z, stable for an entire episode/session, different between arms and between recording sessions) and a representation mismatch (action.*_ee is still the source's own quaternion [qw,qx,qy,qz,x,y,z], not rotmat6) — convert one side to match the other, then anchor action.*_ee's frame 0 to this dataset's frame 0 before comparing the two streams, the same way demo_joint_space anchors it before solving IK.

What's NOT included

• Videos. This repo has no videos/ directory. The source dataset's videos still apply unchanged per episode_index — pair this repo's data/ with adityx23/icl-demo-dataset's videos/ if you need both.