docs/report.tex

% Technical report in ICLR paper style (self-contained; no conference .sty required).
% Compile from this directory:
%   pdflatex report.tex && pdflatex report.tex
\documentclass[11pt,a4paper]{article}

\usepackage[margin=1in]{geometry}
\usepackage{times}
\usepackage{amsmath,amssymb,amsthm}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{tabularx}
\usepackage{hyperref}
\usepackage{xcolor}
\usepackage{microtype}
\usepackage{enumitem}
\usepackage[numbers,sort&compress]{natbib}
\usepackage{caption}
\usepackage{subcaption}

\graphicspath{{figures/}}

\hypersetup{
  colorlinks=true,
  linkcolor=blue!50!black,
  citecolor=blue!50!black,
  urlcolor=blue!50!black
}

\title{\textbf{From RGB Tabletop Demonstrations to a Standing Humanoid:\\
Reconstruction, Retargeting, Frozen Whole-Body Control,\\
and a Path to Dexterous Hand Policies}}

\author{
  Anonymous Authors\\
  \small Technical report, project \texttt{v2d}
}
\date{}

\begin{document}
\maketitle

\begin{abstract}
Learning dexterous tabletop manipulation on a standing humanoid from a single RGB video is blocked by three mismatches that are easy to hide in kinematic replay and expensive to discover in physics.
First, the demonstration is a close-up human hand, not a 29-DoF floating-base robot: objects in DexYCB clips sit about a metre in front of a standing Unitree G1 pelvis, outside the reachable workspace of a frozen whole-body controller.
Second, reconstructed assets are not simulation-ready: SAM~3D meshes are $Y$-up, generated URDFs apply no axis correction, and stock five-finger USDs fail to instantiate under PhysX.
Third, a locomotion-scale policy is not an end-effector oracle, and it is not a grasp policy.
We present a complete video-to-simulation stack that makes these mismatches measurable.
An RGB pipeline (SAM~3 video masks, SAM~3D meshes, HaWoR MANO hands, TAPIR-guided 6-DoF object tracking) is retargeted from MANO to a five-finger robot hand and packed into Isaac Lab / Isaac Sim.
Closed-loop play uses a Unitree G1 with Wuji hands: a frozen SONIC~v1.1 whole-body policy owns \emph{balance} and the 29 body joints; a high-level action is a pelvis-frame wrist target plus a scalar finger closure.
We show that SONIC wrist tracking is repeatable to millimetres but biased by several centimetres, that the bias saturates outside a compact band, and that batched ONNX inference on one H100 is fast enough to sit inside an RL inner loop.
Closed-loop play confirms a standing humanoid whose arm can approach a reconstructed object; the fingers do not yet execute a reliable grasp.
Training a dedicated hand policy---so the robot can lift objects of diverse shape and physics---is the intended next experiment, not a result of this report.
\end{abstract}

\section{Introduction}
\label{sec:intro}

A short RGB clip of a human picking up a can is a tempting supervision signal for a humanoid.
It is also the wrong geometry, the wrong embodiment, and the wrong dynamics for a standing G1.
DexYCB~\citep{chao2021dexycb} records a tabletop right hand from a third-person camera; the reconstructed object in the clips we use lies $\approx 1.1\,\mathrm{m}$ in front of where a standing G1 pelvis would be, and the demonstrated wrist path spans roughly $0.59$--$0.93\,\mathrm{m}$ in $x$.
A standing G1 arm is about $0.6\,\mathrm{m}$ from shoulder to wrist.
Replaying the capture in the robot's world frame therefore asks the controller to reach a location no standing pose can attain.
Kinematic teleport of a floating hand conceals this; contact-rich physics does not.

The second mismatch is assets.
Foundation reconstruction (SAM~3, SAM~3D, HaWoR) produces textured meshes and MANO trajectories that overlay well on the video (Figures~\ref{fig:clips}--\ref{fig:pose}).
Those meshes are $Y$-up; Isaac Lab is $Z$-up; the exported URDF copies the OBJ with identity $\mathrm{rpy}$.
Spawning at identity rotation therefore lays a can on its side, where it has a few millimetres of support margin and rolls at first contact.
The five-finger Inspire USD shipped for G1 is authored as a fixed-base manipulation rig with PhysX mimic joints that this simulator cannot resolve.

The third mismatch is the controller.
GEAR-SONIC~v1.1~\citep{gr00t_wbc} is a whole-body policy for a 29-DoF G1: it balances, tracks a VR-style 3-point target, and must not be confused with the locomotion-only \texttt{G1\_MINIMAL} USD.
It is not a Cartesian IK solver, and it is not a dexterous grasp controller.
Commanding a wrist pose yields a \emph{biased} achieved pose: we measure $1.3\,\mathrm{mm}$ median scatter over a settling window (repeatable) but $8\,\mathrm{cm}$ median offset on a 27-point grid, growing to $>10\,\mathrm{cm}$ when the target crosses the midline or exceeds about $0.45\,\mathrm{m}$ forward of the pelvis.
Fingers are outside SONIC's 29-DoF output entirely.
A high-level policy that treats the wrist command as ground truth will grasp air; a pipeline that never trains the hand will never lift.

This work treats those three facts as the design constraints of a video-to-humanoid pipeline, rather than as bugs to be patched after training.
The contributions are:

\begin{enumerate}[leftmargin=1.4em,itemsep=0.25em]
  \item A self-contained RGB reconstruction stack---SAM~3 instance tracks, SAM~3D object meshes, HaWoR hands, TAPIR-guided 6-DoF object pose---run on DexYCB third-person clips (\texttt{coffee\_can}, \texttt{sugar\_box}), not on egocentric in-house video.
  \item MANO-to-G1-Inspire retargeting (Pink IK, five fingers) and a $Z$-up Isaac Lab pack that preserves the SAM~3D texture instead of replacing it with a procedural primitive.
  \item Closed-loop play of a standing Unitree G1 with Wuji five-finger hands in Isaac Sim, using frozen SONIC on the 29 body joints (name-identical on the Wuji USD) while fingers remain a separate channel.
  \item A wrist-tracking probe that bounds the reachable workspace, and a re-authored table layout in which the object sits inside that band rather than at the capture centre.
  \item A manager-based RL environment in which the learned action is wrist position plus grip, SONIC is an inner-loop \emph{balance} controller, and batched ONNX inference is numerically equivalent to the single-env path at $644\mathrm{k}$ env-steps/s (batch $4096$, H100).
\end{enumerate}

We do \emph{not} report a trained pick or grasp policy.
The hierarchical environment exists; SONIC keeps the robot standing; the clip-relative object-motion objective and a dedicated hand policy are specified in \S\ref{sec:future} and have not been trained.

\section{Related work}
\label{sec:related}

\paragraph{Video to robot data.}
NVIDIA \texttt{video\_to\_data} reconstructs egocentric hand--object motion and retargets MANO to Dex3 for kinematic replay in Isaac Lab.
Our pipeline follows that layout---isolated reconstruction, retarget, and simulation trees---but targets a \emph{standing} 29-DoF G1 with five-finger hands, third-person DexYCB RGB, and a frozen whole-body balancer rather than a floating Dex3 hand.
Kinematic replay remains a debugging tool; it is not the training environment.

\paragraph{Hand--object reconstruction.}
DexYCB~\citep{chao2021dexycb} provides calibrated RGB-D of tabletop grasps; we use only RGB, as would be available from an unstructured demo.
SAM~3~\citep{ravi2024sam2} (and the SAM~3.1 multiplex checkpoint) tracks object and hand masks; SAM~3D lifts the object mask to a textured mesh.
HaWoR~\citep{zhang2025hawor} fits MANO~\citep{romero2017embodied} in a static-camera setting.
TAPIR~\citep{doersch2023tapir} supplies dense 2D tracks that regularise 6-DoF object pose.
MoGe~\citep{wang2025moge} and GeoCalib~\citep{veicht2024geocalib} give metric point maps and gravity so the mesh can be placed in a camera frame with a known up axis.
InterFormer~\citep{lin2026interformer} is available as an egocentric hand--object parser; DexYCB is third-person, so we treat it as optional rather than as the mask source.

\paragraph{Humanoid whole-body control.}
SONIC~\citep{gr00t_wbc} is a VR-teleop whole-body policy: encoder tokens from proprioception and 3-point targets, decoder joint actions at $50\,\mathrm{Hz}$.
Prior humanoid manipulation often either (i)~fixes the base and solves arms with IK, or (ii)~learns locomotion and manipulation jointly in a huge action space.
We freeze SONIC for balance and body joints, and restrict \emph{future} learning to the hand (and a compact wrist command), following the hierarchical pattern of using a low-level stabilizer under a task policy~\citep{peng2018deepmimic,rudin2022rslrl,schulman2017ppo}.

\paragraph{Retargeting.}
Pink / Pinocchio inverse kinematics~\citep{carpentier2019pinocchio} is the standard MANO-to-robot-hand map in recent video-to-robot work.
Dex3 (three fingers) is the NVIDIA default; we retarget to Inspire DFQ (thumb, index, middle, ring, pinky) so the simulated hand matches the five-finger USD used at play time.

\section{Method}
\label{sec:method}

The system is three processes with disjoint Python environments (reconstruction pins NumPy~1.26 for SAM~3; retarget needs Pinocchio; simulation is Isaac Lab on Python~3.12).
They communicate only through files: a reconstruction run directory, a retarget NPZ, and an Isaac Lab pack (Figure~\ref{fig:pipeline} is implicit in this layout).

\subsection{Reconstruction from RGB}
\label{sec:recon}

Given a DexYCB camera folder or a generic MP4, we extract a clip and a reference frame $t_0$ (Figure~\ref{fig:clips}).
Object SAM~3 is prompted with a 2D box (two corners plus a centre point), never with a class name: open-vocabulary text failed on this tabletop (``mug'' does not fire on a DexYCB can).
The hand is prompted with the text ``right hand'' / ``left hand''.
Empty all-black masks are not treated as success, so a prompt frame in which the hand is off-screen does not freeze a broken track (Figure~\ref{fig:sam3}).

SAM~3D lifts the object mask at $t_0$ to a triangle mesh with an MTL/texture.
MoGe predicts a metric point map and focal length; GeoCalib predicts per-frame gravity.
HaWoR, given that focal length and a static-camera flag, writes \texttt{all\_hand\_meshes.npz} (MANO joints, vertices, global rotation) in OpenCV camera coordinates $(x\text{ right},\, y\text{ down},\, z\text{ forward})$.

Object 6-DoF pose is not PnP on a CAD model---we do not assume a YCB mesh ID at runtime.
Fast-SAM3D-style guided pose prediction samples orientations, scores them by render-IoU against the SAM~3 mask, chains poses across time, and is regularised by TAPIR 2D tracks (Figure~\ref{fig:tapir}).
A translation-and-scale optimisation then aligns the mesh in camera frame to the HaWoR hand, producing \texttt{layout\_camera\_frame\_optimized.json} (Figure~\ref{fig:pose}).
Two DexYCB takes are used throughout: \texttt{20200709\_141754} (\texttt{coffee\_can}) and \texttt{20200709\_142553} (\texttt{sugar\_box}), camera serial \texttt{836212060125}.

\subsection{Retargeting MANO to G1 Inspire}
\label{sec:retarget}

HaWoR outputs a MANO wrist and fingertips.
A Pink IK chain, separate from the reconstruction venv, maps those onto the Unitree Inspire DFQ URDF: free-flyer $q_{0:7}$ (wrist pose in camera frame) plus finger joints.
The NVIDIA Dex3 retarget is deliberately not reused; Dex3 has three fingers and different joint names from the Inspire USD.
Figure~\ref{fig:retarget} overlays green MANO on gold Inspire on the same RGB frame.

\subsection{Packing for Isaac Lab}
\label{sec:pack}

Camera-frame poses are mapped to Isaac $Z$-up by
\begin{equation}
  R_{\mathrm{cam}\to z} =
  \begin{pmatrix}
    1 & 0 & 0 \\
    0 & 0 & 1 \\
    0 & -1 & 0
  \end{pmatrix},
  \qquad
  p_z = R_{\mathrm{cam}\to z}\, p_{\mathrm{cam}}.
\end{equation}
Quaternions in the NPZ are stored \textbf{wxyz}; Isaac Lab 6.1 in this tree reports root poses \textbf{xyzw}.
Confusing the two at the policy input is a $180^\circ$ yaw and looks like a broken model, not a crash.

The object URDF references the SAM~3D OBJ at the optimised scale and copies the MTL and \texttt{map\_Kd} texture.
The spawn config must not set a \texttt{visual\_material}: Isaac's URDF converter then replaces the texture with a flat preview colour (the green cuboid in early kinematic replays, Figure~\ref{fig:kinematic}).

Collision is a convex hull.
A can and a sugar box are convex enough that convex decomposition only costs startup time.

Rest orientation is taken from the demonstration: the clip quaternion at the frame of lowest object $z$, with residual tilt snapped so the body axis already nearest world $+Z$ becomes exactly vertical.
That axis is not a global mesh convention (the coffee can rests on body $+Y$; the sugar box in this take rests on $-Y$ and is a $5\,\mathrm{cm}$ slab).
Support height is minus the lowest rotated vertex, not half the axis-aligned bounding box.

\subsection{Frozen SONIC as the body balancer}
\label{sec:sonic}

SONIC~v1.1 is an ONNX encoder--decoder pair.
The encoder sees a 1751-D vector: joint positions and velocities (Isaac Lab order, not MuJoCo order), angular velocity, projected gravity, a 6D heading relative to a reference pelvis, a one-hot encoder mode, and VR 3-point targets (left wrist, right wrist, torso) expressed in the \emph{reference} pelvis, not the robot pelvis.
The third point is \texttt{torso\_link} offset by $+0.35\,\mathrm{m}$, not a head body.
History is oldest-first.
The 6D rotation is the first two \emph{rows} of the rotation matrix flattened, $[m_{00}, m_{01}, m_{10}, m_{11}, m_{20}, m_{21}]$; identity is $[1,0,0,1,0,0]$.

We use \textbf{teleop} mode: the lower body tracks a standing reference; the right-wrist target is the reconstructed (then relocated) hand root; left wrist and torso hold the standing FK.
Hand orientation is held at the standing wrist quaternion.
The DexYCB hand frame is MANO, not \texttt{right\_wrist\_yaw\_link}; tracking that rotation drives the wrist into its limits.

SONIC emits 29 joint position offsets.
Fingers are not in that set.
On the Dex3 USD there are 43 joints; on the patched Inspire USD, 53 plus the free root; on the Wuji USD (\texttt{g1\_wuji\_no\_merge.usd}) there are 69 (29 body + 20 DoF per hand).
Anything that reads or writes joints must resolve SONIC's 29 \emph{by name}.
Indexing $0..28$ silently drives the wrong joints the moment a hand USD is swapped.

\paragraph{G1 with Wuji hands.}
Closed-loop Isaac Sim play uses the Unitree G1 body with the Wuji five-finger end-effector (Figure~\ref{fig:wuji}).
The Wuji USD carries all 29 SONIC body joints under the same names as the Inspire and Dex3 assets, so the balancer is unchanged; only the finger set differs (\texttt{\{left,right\}\_finger$\langle$1--5$\rangle$\_joint$\langle$1--4$\rangle$}, no PhysX mimic joints).
That is the embodiment we train toward: a standing G1 whose hands can, in principle, wrap or pinch reconstructed objects. SONIC still does not drive those fingers.

This is the capability we actually have: a policy that can keep a floating-base G1 standing while the arm tracks a wrist target.
It does not close a grasp, does not reason about contact, and does not vary finger shape with object geometry.
That gap is the subject of \S\ref{sec:future}.

\paragraph{Inspire USD.}
The stock \texttt{g1\_29dof\_inspire\_hand.usd} does not spawn as a floating-base articulation.
\texttt{PhysxMimicJointAPI} cannot find its reference joints, and \texttt{PhysicsArticulationRootAPI} sits on a \texttt{PhysicsFixedJoint} at \texttt{root\_joint}.
Disabling that joint to free the base removes the articulation root with it.
We mirror the asset, strip both the mimic properties \emph{and} the applied API schema (leaving the schema produces ``must have exactly 1 \texttt{referenceJoint}'' once per joint per env), and move the articulation root onto the pelvis, matching stock \texttt{g1.usd}.
The six coupled joints per hand then become independently actuated; linkage gearing can be reimposed in software if needed, with a sign check against joint-limit intervals that are not mirrors of each other.

\paragraph{Batching.}
The shipped ONNX graphs are traced at batch 1.
Twenty \texttt{Reshape} nodes spell the batch as literal $1$; we rewrite them to $-1$.
The encoder-mode one-hot is a \texttt{ScatterND} into a constant of shape $[1,3]$ indexed by a \texttt{arange} that constant-folded to $[0]$, so every environment past the first would receive a zero token.
We replace it with $\mathbf{1}[b,k] = \mathbb{I}[\textit{encoder\_index}(b)=k]$.
CUDA execution requires preloading CUDA~13 \texttt{libcudart}/\texttt{cublas}/\texttt{cublasLt}/\texttt{curand} with \texttt{RTLD\_LOCAL}: \texttt{LD\_LIBRARY\_PATH} would shadow the CUDA~12 libraries PyTorch is built against.

On a standing reference, almost the entire 1751-D encoder is constant across environments.
\texttt{BatchedSonicController} builds it once and rewrites only heading (6) and VR 3-point targets (21) per tick.
Packing is bit-exact against the single-env path.

\subsection{Workspace re-authoring}
\label{sec:workspace}

Let $p^{\mathrm{cmd}}$ be a commanded right-wrist position in the pelvis frame and $p^{\mathrm{ach}}$ the FK position after SONIC has settled.
The probe in \S\ref{sec:probe} shows $\|p^{\mathrm{ach}}-p^{\mathrm{cmd}}\|$ is a smooth function of $p^{\mathrm{cmd}}$, small in a band
\begin{equation}
  x \in [0.30, 0.42],\quad
  y \in [-0.28, -0.12],\quad
  z \in [-0.05, 0.15]
\end{equation}
(metres, pelvis frame), and saturating outside it.
The DexYCB object at capture centre is not in that band.

We therefore \emph{do not} replay the capture layout.
The robot pelvis is at the world origin, table top at $z=0.75\,\mathrm{m}$, table extent $(0.80, 1.00, 0.05)\,\mathrm{m}$ centred at $x=0.60$ (a floating slab: legs at this depth collide with the feet).
The object's first frame is translated so its $xy$ lands at $(0.36, -0.20)$, the centre of the well-tracked band.
The reconstructed hand is translated by the same $\Delta xy$, so grasp-relative geometry is preserved; only the scene origin moves.

Reset yaw is composed on the \emph{left}, $q_{\mathrm{yaw}}^{\mathrm{world}} q_{\mathrm{rest}}$.
Isaac Lab's stock \texttt{reset\_root\_state\_uniform} composes on the right, so its ``yaw'' is about body $z$.
After a $90^\circ$ rest roll, body $z$ is horizontal and that term tips the object over.

\subsection{Hierarchical MDP}
\label{sec:mdp}

The Gym task \texttt{V2D-G1-SonicManip-v0} is a manager-based Isaac Lab environment.
Physics runs at $200\,\mathrm{Hz}$, SONIC at $50\,\mathrm{Hz}$, the policy at $25\,\mathrm{Hz}$ (decimation $8$, a multiple of the SONIC stride).

\textbf{Action.} $a = (a_x, a_y, a_z, a_g) \in [-1,1]^4$.
The first three coordinates are affinely mapped into the probe band and become the right-wrist VR target; $a_g$ interpolates finger joints from open to closed.
Wrist orientation is held at the standing pose.
This 4-D interface is a scaffold for the hand policy in \S\ref{sec:future}, not a claim that one grip scalar is sufficient.

\textbf{Observation.} Projected gravity, \emph{achieved} right-wrist position in the pelvis frame, object position, object-to-wrist vector, object linear velocity, grip, last action.
The policy never sees a joint.
Achieved wrist pose is required because of the bias in \S\ref{sec:probe}: commanding $p^{\mathrm{cmd}}$ and observing $p^{\mathrm{cmd}}$ again would hide several centimetres of systematic error.

\textbf{Placeholder rewards} currently used for environment smoke tests are reach (tanh of wrist--object distance) and lift relative to rest height, gated on proximity, plus alive / fall / action-rate / torso-upright terms.
Lift is measured from the object origin's resting $z$, not from the table top: the origin is the mesh centroid, already $\sim 7\,\mathrm{cm}$ up when the can is untouched, so a table-relative lift reward pays out for doing nothing.

\section{Experiments}
\label{sec:expts}

All simulation numbers are from Isaac Lab / Isaac Sim on a single H100 node unless noted.
Reconstruction and retargeting are qualitative plus the geometric measurements below.
Stills in this section are exported from the corresponding run videos under \texttt{reconstruction/runs/} and \texttt{simulation/runs/}.

\subsection{Reconstruction and retargeting}
\label{sec:exp-recon}

\begin{figure}[t]
  \centering
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{clip_can.png}
    \caption{Coffee can (\texttt{141754}), reference frame.}
  \end{subfigure}
  \hfill
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{clip_box.png}
    \caption{Sugar box (\texttt{142553}), reference frame.}
  \end{subfigure}
  \caption{Two DexYCB RGB clips used as demonstrations. Object SAM~3 is box-prompted; the hand is text-prompted. Meshes and 6-DoF tracks are lifted from these videos, not from YCB CAD.}
  \label{fig:clips}
\end{figure}

\begin{figure}[t]
  \centering
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{sam3_object.png}
    \caption{SAM~3 object track (box prompt) on the coffee can.}
  \end{subfigure}
  \hfill
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{sam3_hand.png}
    \caption{SAM~3 right-hand track (text prompt) at grasp onset.}
  \end{subfigure}
  \caption{Instance masks that drive the rest of reconstruction. Open-vocabulary class names failed on this tabletop; geometric and text prompts do not.}
  \label{fig:sam3}
\end{figure}

\begin{figure}[t]
  \centering
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{tapir_pairs.png}
    \caption{TAPIR correspondences between consecutive frames (19 matches shown).}
  \end{subfigure}
  \hfill
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{tapir_motion.png}
    \caption{Translation speed, rotation velocity, and visible-point count over the clip.}
  \end{subfigure}
  \caption{TAPIR-guided object motion. The can is nearly stationary until frame~$\sim$30, then lifts (peak $\approx 10\,\mathrm{px}/\mathrm{frame}$); 20 tracks remain visible until the last frame. This motion, after $Z$-up packing and workspace translation, is the intended object-tracking target for a future hand policy.}
  \label{fig:tapir}
\end{figure}

\begin{figure}[t]
  \centering
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{sam3d_projected.png}
    \caption{SAM~3D mesh projected into the RGB frame.}
  \end{subfigure}
  \hfill
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{sam3d_opt.png}
    \caption{Hand--object layout after scale/translation optimisation (red~=~optimised object, cyan~=~HaWoR vertices).}
  \end{subfigure}
  \caption{6-DoF object pose from SAM~3D, not from a YCB CAD model. The optimisation reduces a layout residual from $0.095$ to $0.009$ on the illustrated frame.}
  \label{fig:pose}
\end{figure}

\begin{figure}[t]
  \centering
  \includegraphics[width=0.82\linewidth]{retarget_overlay.jpg}
  \caption{MANO-to-G1-Inspire retarget overlay on DexYCB RGB (green~=~MANO, gold~=~Inspire). Finger names and DoF match the five-finger USD used in Isaac Sim, not Dex3.}
  \label{fig:retarget}
\end{figure}

Both clips produce SAM~3 tracks, a textured SAM~3D mesh, HaWoR MANO, and an optimised camera-frame object trajectory.
Offline retarget overlays (Figure~\ref{fig:retarget}) and kinematic Isaac Lab replay (Figure~\ref{fig:kinematic}) confirm that the pack is self-consistent in $Z$-up.
Those videos do not test balance, contact, or reachability: the hand is teleported.

\subsection{SONIC closed-loop play in Isaac Sim}
\label{sec:exp-sonic}

\begin{figure}[t]
  \centering
  \begin{subfigure}{0.32\textwidth}
    \includegraphics[width=\linewidth]{sim_sonic_stand.jpg}
    \caption{Standing reference (\texttt{g1} mode).}
  \end{subfigure}
  \hfill
  \begin{subfigure}{0.32\textwidth}
    \includegraphics[width=\linewidth]{sim_sonic_replay_can.jpg}
    \caption{Teleop on the coffee-can pack.}
  \end{subfigure}
  \hfill
  \begin{subfigure}{0.32\textwidth}
    \includegraphics[width=\linewidth]{sim_sonic_replay_box.jpg}
    \caption{Teleop on the sugar-box pack.}
  \end{subfigure}
  \caption{Isaac Sim closed-loop play with frozen SONIC. The robot holds a standing pelvis ($\approx 0.78\,\mathrm{m}$) while the right wrist tracks the relocated demonstration. The object is the reconstructed SAM~3D mesh, not a primitive. Fingers are open: SONIC does not drive them.}
  \label{fig:sim}
\end{figure}

\begin{figure}[t]
  \centering
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{sim_kinematic_late.jpg}
    \caption{Kinematic pack replay (floating hand; texture bug shown as a green cylinder).}
  \end{subfigure}
  \hfill
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{sim_sonic_manip.jpg}
    \caption{\texttt{V2D-G1-SonicManip}: reachable table, reconstructed object, SONIC-in-the-loop.}
  \end{subfigure}
  \caption{Left: kinematic debugging of the Isaac pack (no balance). Right: the hierarchical environment in which a future hand policy would act. The robot reaches the object; a scripted closure does not yet lift it.}
  \label{fig:kinematic}
\end{figure}

On a standing reference (\texttt{--mode g1}), the robot holds a pelvis height of $\approx 0.78\,\mathrm{m}$ for a full episode; $\max|a| < 0.5$ is the health check that joint order and 6D flattening are correct (Figure~\ref{fig:sim}a).
Values near $4$--$5$ indicate one of those conventions is swapped.

In teleop mode the body remains upright while the right arm tracks the relocated wrist target (Figures~\ref{fig:sim}b--c).
Minimum root height over smoke rollouts stayed above $0.75\,\mathrm{m}$ (fall threshold $0.5\,\mathrm{m}$).

The play embodiment is a Unitree G1 with Wuji five-finger hands (Figure~\ref{fig:wuji}): the robot approaches a reconstructed object on a table while SONIC keeps it standing.
Fingers are visible and independently actuated, but they are not yet a trained grasp policy---the hand reaches; it does not lift.
This is the result we actually have: a standing G1--Wuji whose arm moves, not a completed grasp.

\begin{figure}[t]
  \centering
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{sim_wuji_reach.jpg}
    \caption{Approach: G1--Wuji reaching the reconstructed can.}
  \end{subfigure}
  \hfill
  \begin{subfigure}{0.48\textwidth}
    \includegraphics[width=\linewidth]{sim_wuji_grasp.jpg}
    \caption{Near-contact: right Wuji hand open above the object.}
  \end{subfigure}
  \caption{Isaac Sim closed-loop play on the Unitree G1 with Wuji five-finger hands (20 DoF per hand). The whole-body policy balances the floating base; the Wuji hand is the end-effector we intend a future grasp policy to drive.}
  \label{fig:wuji}
\end{figure}

\subsection{Wrist tracking probe}
\label{sec:probe}

\begin{table}[t]
  \centering
  \caption{SONIC right-wrist tracking in the pelvis frame (27-point grid, median over a settling window). Scatter is $1.3\,\mathrm{mm}$; the offset is structured, not noise.}
  \label{tab:probe}
  \begin{tabular}{lrr}
    \toprule
    commanded & median $\|e\|$ (m) & signed bias on that axis (m) \\
    \midrule
    $x = 0.40$ & $0.042$ & $+0.013$ \\
    $x = 0.55$ & $0.101$ & $-0.055$ (out of reach) \\
    $y = -0.30$ & $0.035$ & $+0.028$ \\
    $y = 0.00$  & $0.116$ & $+0.062$ (cannot cross midline) \\
    $z = -0.05$ & $0.032$ & $+0.026$ \\
    $z = +0.25$ & $0.105$ & $+0.077$ \\
    \bottomrule
  \end{tabular}
\end{table}

Table~\ref{tab:probe} is the measurement that decides the rest of the system.
Repeatability ($1.3\,\mathrm{mm}$) means a high-level policy can close a residual with feedback.
An $8\,\mathrm{cm}$ median bias that grows outside a compact band means (i)~the action box must be that band, not the kinematic limit, and (ii)~observations must include achieved pose.
A target at $(x,z)=(0.45, 0.03)$ is $0.59\,\mathrm{m}$ from the shoulder---the length of the arm---which matches the saturation we see at $x=0.55$.

The DexYCB can at $1.10\,\mathrm{m}$ in front of the pelvis is unreachable.
Even standing flush against a $1.2\,\mathrm{m}$-deep table with the object at its centre leaves $0.60\,\mathrm{m}$ of depth.
Use the clip for relative grasp geometry; author the scene for the robot.

\subsection{Object rest pose}
\label{sec:exp-rest}

For the coffee-can mesh (URDF scale $0.139$), identity rotation in $Z$-up world yields $198$ contact vertices and $1.2\,\mathrm{cm}$ of support-polygon margin: technically balanced, practically the first touch rolls it.
The clip quaternion is $118.5^\circ$ from identity; flattening residual tilt yields $2066$ contact vertices and $6.1\,\mathrm{cm}$ of margin, origin $6.94\,\mathrm{cm}$ above the table, footprint $\approx 12.2\times 12.9\,\mathrm{cm}$, height $13.9\,\mathrm{cm}$.
The sugar-box mesh (scale $0.191$), using the demonstration's resting face rather than a hardcoded $+Y$ up-axis, sits as a $5.1\,\mathrm{cm}$ slab with a $15$--$19\,\mathrm{cm}$ footprint---the pose the human actually used, not the most ``canonical'' YCB orientation.

\subsection{Batched inference}
\label{sec:exp-batch}

\begin{table}[t]
  \centering
  \caption{Encoder+decoder throughput per control tick on one H100. CPU is $\sim 100\times$ too slow to host SONIC inside RL; CUDA is not.}
  \label{tab:batch}
  \begin{tabular}{rrr}
    \toprule
    batch & CPU (env-steps/s) & CUDA (env-steps/s) \\
    \midrule
    $256$  & $6.5\times 10^3$ & $2.04\times 10^5$ \\
    $4096$ & $6.2\times 10^3$ & $6.44\times 10^5$ ($6.4\,\mathrm{ms}$/tick) \\
    \bottomrule
  \end{tabular}
\end{table}

Table~\ref{tab:batch} answers whether a frozen ONNX inner loop is affordable.
At batch $4096$ the controller is $6.4\,\mathrm{ms}$ per tick on GPU; the RL step is $40\,\mathrm{ms}$ of simulated time at $25\,\mathrm{Hz}$, so inference is not the bottleneck.
CPU at $\sim 6\mathrm{k}$ env-steps/s would be.

\subsection{Scripted manipulation smoke}
\label{sec:exp-smoke}

An open-loop reach / close / lift schedule in \texttt{V2D-G1-SonicManip} keeps the robot standing and reduces hand--object distance from $\approx 24\,\mathrm{cm}$ to $\approx 6\,\mathrm{cm}$ (Figure~\ref{fig:kinematic}b).
That residual matches the probe bias; it is larger than a $6\,\mathrm{cm}$-scale can, so a scripted grasp does not lift.
An earlier ``$4/4$ picked up'' metric compared object $z$ to the table top and was identically true at rest.
The honest reading: the environment steps, SONIC balances under a moving arm, and the hand does not yet get close enough to evaluate contact or the clip-relative objective in \S\ref{sec:future}.

\section{Future work: training a hand policy}
\label{sec:future}

The scientific goal is not ``the robot can stand.''
It is: \textbf{SONIC continues to own balance; a learned hand policy closes contact and reproduces reconstructed object motion} after the rigid workspace translation of \S\ref{sec:workspace}, across objects that differ in shape, mass, and friction.

Let $x_t$ be the simulated object pose and $x^{\mathrm{d}}_t$ the demo pose at the corresponding clip phase, both in the re-authored world.
Two residuals make ``match the demonstration'' precise:
\begin{align}
  e^{\mathrm{abs}}_t
  &= (x_t - x_0) - (x^{\mathrm{d}}_t - x^{\mathrm{d}}_0),
  \label{eq:abs} \\
  e^{\mathrm{rel}}_t
  &= (x_t - x_{t-1}) - (x^{\mathrm{d}}_t - x^{\mathrm{d}}_{t-1}).
  \label{eq:rel}
\end{align}
Equation~\eqref{eq:abs} prevents drift of the whole path; \eqref{eq:rel} matches frame-to-frame motion (the TAPIR lift in Figure~\ref{fig:tapir}b after packing).
A reward of the form
\begin{equation}
  r_t
  = \exp\!\big(-\|e^{\mathrm{abs}}_t\|^2_{\Sigma^{-1}}\big)
  + \exp\!\big(-\|e^{\mathrm{rel}}_t\|^2_{\Lambda^{-1}}\big)
  + r^{\mathrm{contact}}_t
  - \lambda \|a_t - a_{t-1}\|^2
\end{equation}
is the natural training signal, with $r^{\mathrm{contact}}$ so the policy cannot score by sliding the object with the table.

\paragraph{Why a separate hand policy.}
SONIC is a 29-DoF body controller trained for VR teleop.
Its action does not include fingers; its observations do not include fingertip forces or object geometry beyond what a wrist target implies.
Retraining the whole body to pick up a can would throw away a working balancer and explode the action space.
The architecture we intend is therefore hierarchical: freeze SONIC, learn on top.

The 4-D wrist-and-grip action in \S\ref{sec:mdp} is the smallest interface that can express a lift.
It is almost certainly too small for diverse shapes: the coffee can wants a wrap around a $13\,\mathrm{cm}$ cylinder; the sugar box in this take is a $5\,\mathrm{cm}$ slab and wants a pinch or side grasp.
The planned growth, in order, is:

\begin{enumerate}[leftmargin=1.4em,itemsep=0.3em]
  \item \textbf{Expert-wrist smoke (no learning).}
  Drive the relocated HaWoR wrist through the existing action box and close fingers on a distance schedule.
  Gate: hand--object distance enters a contact-scale band and $\|e^{\mathrm{rel}}\|$ drops below the ``object never moved'' baseline.
  If the object never leaves rest, stop and fix tracking bias, spawn pose, or finger timing. Do not train.

  \item \textbf{Clip-conditioned MDP.}
  Load \texttt{isaaclab\_replay.npz} as a time-indexed reference; add $e^{\mathrm{abs}}$, $e^{\mathrm{rel}}$, and phase $t/T$ to observations; replace lift-as-main with the reward above; set episode length to clip duration.

  \item \textbf{PPO on the 4-D action, SONIC frozen.}
  Small MLP, RSL-RL~\citep{rudin2022rslrl,schulman2017ppo}, success metrics $\mathbb{E}\|e^{\mathrm{abs}}\|$, $\mathbb{E}\|e^{\mathrm{rel}}\|$ in the last $20\%$ of the episode, fraction of steps with hand--object distance $<4\,\mathrm{cm}$, fall rate.
  Compare against the expert-wrist baseline, not against a random policy.

  \item \textbf{Richer hand actions, only if (3) saturates on grasp geometry.}
  Wrist orientation (3-D / 6-D) if the clip approach is palm-down and standing yaw cannot grasp; then split grip into thumb vs.\ four-finger groups---not 12 independent DoF on day one.

  \item \textbf{Object diversity.}
  Train and evaluate on multiple reconstructed DexYCB packs (can, box, and further exports), randomising mass and friction around the URDF defaults.
  The point of reconstruction is that each clip is a different mesh and a different affordance, not a second texture on the same cylinder.
\end{enumerate}

\paragraph{What we will not do in the next round.}
End-to-end PPO on 29 or 53 joints (SONIC stays frozen).
Replaying DexYCB world coordinates with the object at $1.1\,\mathrm{m}$.
Per-finger imitation of MANO joint angles as the \emph{primary} loss (object motion is the target; fingers are a means).

\section{Limitations}
\label{sec:limits}

\textbf{No trained hand policy.} Hierarchical PPO is wired (RSL-RL, small MLP, 4-D action) and has not been run against \eqref{eq:abs}--\eqref{eq:rel}.

\textbf{Wrist orientation is unused.} Grasps that need a specific approach angle cannot be expressed.
Only position tracking was measured.

\textbf{One grip scalar.} Coupled Inspire joints are independently actuated after mimic stripping; a single closure interpolates all of them.
Per-finger actions are premature until contact exists at all.

\textbf{Single camera, static rig.} HaWoR is run with \texttt{--static\_camera}. DexYCB cameras are calibrated; a handheld demo would need SfM or a different HaWoR mode.

\textbf{Third-person DexYCB.} InterFormer is egocentric SOTA; we did not use it as the primary masker.

\textbf{Convex hull, no deformables.} Fine finger--rim contact on a can may need a better collision approximation than a hull.

\textbf{Two clips.} Coffee can and sugar box from one subject and camera. Nothing here is a dataset paper.

\section{Conclusion}
\label{sec:concl}

A standing humanoid cannot execute a DexYCB tabletop demo by loading the reconstructed trajectory into its world frame.
The object is too far, the mesh is in the wrong up-axis, the five-finger USD does not spawn, and the whole-body controller tracks wrist targets with centimetres of structured bias and does not move the fingers.
We built the pipeline that makes each of those statements a measurement rather than a guess: SAM~3 / SAM~3D / TAPIR reconstruction, MANO-to-Inspire retargeting, and an Isaac Sim task in which SONIC owns balance while a small action space owns the hand.
The next experiment is not more USD repair.
It is to train a hand policy that closes contact and matches clip-relative object motion---the ability the current balancer demonstrably lacks.

\bibliographystyle{plainnat}
\begin{thebibliography}{99}

\bibitem[Carpentier et~al.(2019)]{carpentier2019pinocchio}
J.~Carpentier, G.~Saurel, G.~Buondonno, J.~Mirabel, F.~Lamiraux, O.~Stasse, and N.~Mansard.
The Pinocchio C++ library -- a fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives.
In \emph{IEEE/SICE International Symposium on System Integration (SII)}, 2019.

\bibitem[Chao et~al.(2021)]{chao2021dexycb}
Y.-W.~Chao, W.~Yang, Y.~Xiang, P.~Molchanov, A.~Handa, J.~Tremblay, Y.~S.~Narang, K.~Van~Wyk, U.~Iqbal, S.~Birchfield, J.~Kautz, and D.~Fox.
DexYCB: A benchmark for capturing hand grasping of objects.
In \emph{IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, 2021.

\bibitem[Doersch et~al.(2023)]{doersch2023tapir}
C.~Doersch, Y.~Yang, M.~Vecerik, D.~Gokay, A.~Gupta, Y.~Aytar, J.~Carreira, and A.~Zisserman.
TAPIR: Tracking any point with per-frame initialisation and temporal refinement.
In \emph{IEEE/CVF International Conference on Computer Vision (ICCV)}, 2023.

\bibitem[Lin et~al.(2026)]{lin2026interformer}
Z.~Lin et~al.
Interaction-aware representation modeling with co-occurrence consistency for egocentric hand-object parsing.
In \emph{International Conference on Learning Representations (ICLR)}, 2026.

\bibitem[Zhang et~al.(2025)]{zhang2025hawor}
J.~Zhang et~al.
HaWoR: World-space hand motion reconstruction from videos.
2025.
See the HaWoR module README for the camera-ready citation.

\bibitem[Peng et~al.(2018)]{peng2018deepmimic}
X.~B.~Peng, P.~Abbeel, S.~Levine, and M.~van~de Panne.
DeepMimic: Example-guided deep reinforcement learning of physics-based character skills.
\emph{ACM Transactions on Graphics}, 37(4), 2018.

\bibitem[Ravi et~al.(2024)]{ravi2024sam2}
N.~Ravi, V.~Gabeur, Y.-T.~Hu, R.~Hu, C.~Ryali, T.~Ma, H.~Khedr, R.~R{\"a}dle, C.~Rolland, L.~Gustafson, E.~Mintun, J.~Pan, K.~V.~Alwala, N.~Caron, C.-Y.~Wu, R.~Girshick, P.~Doll{\'a}r, and C.~Feichtenhofer.
SAM 2: Segment anything in images and videos.
\textit{arXiv:2408.00714}, 2024.

\bibitem[Romero et~al.(2017)]{romero2017embodied}
J.~Romero, D.~Tzionas, and M.~J.~Black.
Embodied hands: Modeling and capturing hands and bodies together.
\emph{ACM Transactions on Graphics (Proc.\ SIGGRAPH Asia)}, 36(6), 2017.

\bibitem[Rudin et~al.(2022)]{rudin2022rslrl}
N.~Rudin, D.~Hoeller, P.~Reist, and M.~Hutter.
Learning to walk in minutes using massively parallel deep reinforcement learning.
In \emph{Conference on Robot Learning (CoRL)}, 2022.

\bibitem[Schulman et~al.(2017)]{schulman2017ppo}
J.~Schulman, F.~Wolski, P.~Dhariwal, A.~Radford, and O.~Klimov.
Proximal policy optimization algorithms.
\textit{arXiv:1707.06347}, 2017.

\bibitem[Veicht et~al.(2024)]{veicht2024geocalib}
A.~Veicht, P.-E.~Sarlin, P.~Lindenberger, and M.~Pollefeys.
GeoCalib: Learning single-image calibration with geometric optimization.
In \emph{European Conference on Computer Vision (ECCV)}, 2024.

\bibitem[Wang et~al.(2025)]{wang2025moge}
R.~Wang et~al.
MoGe: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision.
In \emph{IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, 2025.

\bibitem[NVIDIA(2025)]{gr00t_wbc}
NVIDIA GEAR.
GR00T whole-body control / SONIC.
\url{https://github.com/NVlabs/GR00T-WholeBodyControl}, 2025.

\end{thebibliography}

\appendix
\section{Reproducibility sketch}
\label{app:repro}

Reconstruction (Python~3.10 venv):
\begin{verbatim}
cd reconstruction
./run_pipeline.sh --dexycb ../data/dexycb/.../836212060125 \
  --object coffee_can --hand right --box box.json
\end{verbatim}
Retarget and Isaac pack (separate venv):
\begin{verbatim}
cd retarget
./hawor_to_g1.sh --video-dir ../reconstruction/runs/<seq> --hands auto
./export_isaaclab.sh --video-dir ../reconstruction/runs/<seq>
\end{verbatim}
Closed-loop SONIC (Isaac Lab venv, GPU node):
\begin{verbatim}
cd simulation
./replay_sonic.sh --headless --hands inspire \
  --npz ../reconstruction/runs/<seq>/obj_tracking_out/isaaclab_replay.npz
\end{verbatim}
Wrist probe and batched export: \texttt{jobs/sonic\_probe.sh}, \texttt{scripts/export\_sonic\_dynamic\_batch.py}.
Hierarchical env: Gym id \texttt{V2D-G1-SonicManip-v0}; smoke \texttt{jobs/sonic\_manip\_smoke.sh}.

\section{Convention checklist}
\label{app:conv}

If SONIC thrashes on a standing reference, check in this order: (1)~joint order is Isaac Lab names, not MuJoCo \texttt{default\_angles} order; (2)~6D identity is $[1,0,0,1,0,0]$; (3)~root quaternion from the simulator is converted xyzw$\to$wxyz before the encoder; (4)~VR targets are in the reference pelvis; (5)~history is oldest-first.

\section{Figure sources}
\label{app:figs}

Stills under \texttt{docs/figures/} are copied or extracted from:
\texttt{reconstruction/runs/20200709\_141754\_836212060125} (SAM~3 overlays, TAPIR pairs, SAM~3D projected/optimised layouts, retarget overlay, kinematic replay, SONIC replay of the can)
and
\texttt{simulation/runs/} (standing smoke, teleop smoke, \texttt{sonic\_manip\_smoke.mp4}, G1--Wuji play stills from \texttt{rl-video-step-0.mp4}).
The sugar-box RGB frame and SONIC replay are from \texttt{reconstruction/runs/20200709\_142553\_836212060125}.

\end{document}