WARP: Whole-body Retargeting from Offline Human Motion via Closed-Form c-SEW

TL;DR

WARP retargets offline human motion capture into robot actions for whole-body mobile manipulation via a closed-form Shoulder–Elbow–Wrist (c-SEW) solver that treats palm pose as a hard constraint and uses the remaining redundancy to reproduce the human’s arm geometry. Because the target setting is offline — no operator in the loop to absorb tracking error — retargeting quality is the supervision, and the design trades expressive optimization objectives for exact-by-construction end-effector matching. Evaluated on 514 BONES-SEED clips across RB-Y1, dual Kinova3, and OpenArm, WARP cuts palm error >150× vs the MINK-EF optimization IK baseline, hits machine-precision orientation error, and produces the most self-collision- and joint-limit-feasible trajectories. Positioned as the enabling substrate for training whole-body mobile-manipulation policies from offline human demos alone.

Key claims

Offline retargeting requires exact end-effector matching because the retargeted trajectory replays open-loop with no operator to absorb error — soft-weighted optimization objectives (used by MINK-EF/MINK-TE, SEW-mimic) leave residual palm error that compounds at replay [§Method / project page].
c-SEW gives a closed-form solution: fix palm pose as a hard constraint, then use the remaining kinematic redundancy in the shoulder–elbow–wrist chain to reproduce the human’s arm geometry, yielding exactly one robot configuration per human pose in microseconds [§Method].
Precision result: on 514 BONES-SEED clips, WARP achieves 0.0046 mm mean palm error and 0.046 mm P95 vs MINK-EF’s 0.701 mm and 1.853 mm — >150× reduction — at machine-precision orientation error (8.74e-6 deg vs 0.0107 deg) [Table, offline column].
Feasibility result: WARP produces the lowest self-collision fraction (0.0047 vs MINK-EF’s 0.1610) and lowest joint-limit-violation fraction (0.163 vs MINK-EF’s 0.977), showing that human pose is itself a strong feasibility prior when used as the geometric anchor rather than as a soft cost [Table].
Consistency result: WARP’s deterministic closed-form solver eliminates the near-identical-input-different-output pathology of optimization IK on redundant humanoids (RMS deg 6.66e-14 vs MINK-EF’s 3.117), which the authors argue makes the resulting demo dataset more learnable [Table, solver variation columns].
Whole-body-intent preservation: end-effector-only retargeting discards torso, elbow, and base motion — WARP retains those channels so downstream policies can learn the coordinated whole-body strategy the human used [§Findings 3].
Cross-embodiment transfer: the same c-SEW recipe applies to RB-Y1 humanoid, dual-Kinova3 bimanual, and OpenArm — and the authors report it gives the first whole-body mobile-manipulation policy trained from offline human demos alone [§Results].

Method

WARP is a closed-form inverse-kinematics solver for retargeting human motion capture onto redundant humanoid and bimanual robot arms. The core primitive is c-SEW: a closed-form Shoulder–Elbow–Wrist analytical IK derived from SEW-mimic, in which the palm pose (position + orientation) is imposed as a hard constraint that must be matched exactly. The residual degrees of freedom in the arm’s kinematic chain are then used to reproduce the geometric configuration (shoulder, elbow, wrist angles) of the human’s arm — so every human pose maps to exactly one robot configuration, computed in microseconds per frame with no iterative solver, no cost weighting, and no reward tuning. Whole-body components (torso, base) are propagated through parallel channels so the retargeted trajectory preserves the human’s coordinated whole-body solution rather than reducing to end-effector-only playback. Because the target regime is offline (mobile teleop is expensive/slow/unnatural and requires a robot in the loop for every demo), the retargeted trajectory has to replay open-loop and the retargeted trajectory is the supervision — motivating the exactness-over-flexibility design choice.

Results

Evaluated on 514 BONES-SEED clips. In the offline regime, WARP achieves 0.0046 mm mean palm tracking error vs MINK-EF 0.701 (>150× improvement), SEW-mimic 178.979, and MINK-TE 18.557; P95 palm error 0.046 mm vs MINK-EF 1.853. Orientation error 8.74e-6 deg (machine precision) vs MINK-EF 0.0107. Self-collision fraction 0.0047 vs MINK-EF 0.1610 vs MINK-TE 0.0852 (best on the metric). Joint-limit-violation fraction 0.163 vs MINK-EF 0.977 (best on the metric). Solver-variation metrics (NNAD, PCA eig, RMS deg) are at machine precision (1.14e-25 / 6.66e-14 / near-zero) vs MINK-EF’s 173.49 / 3.117 — orders of magnitude more consistent, matching only the SEW-M baseline which is orders of magnitude worse on tracking error. Solve speed ~30× faster than baselines (an hour vs a day for the full SEED corpus). Cross-embodiment demonstrations reported on RB-Y1 humanoid, dual Kinova3, and OpenArm. The authors claim WARP gives the whole-body mobile-manipulation policy trained from offline human demos alone — no in-scene teleop.

Why it’s interesting

WARP stakes out a tenth position on the interface-location axis tracked by Human-to-Robot Retargeting: rather than choosing where the shared human–robot interface lives (reduced-DoF action space, 4D reconstruction, task intent, wrist translation, contact topology, hardware-locked collection tool, …), WARP argues the axis is misframed and the real lever is solver class — a closed-form analytical IK with hard end-effector constraints beats every optimization-IK recipe on the same interface (full arm kinematics) because the offline regime punishes soft residual error. This is a direct methodological contrast with ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting (physics-in-loop bilevel optimization) which pays for the same end-goal — hardware-executable trajectories — with an outer-loop RL cost, and with Do as I Do: Dexterous Manipulation Data from Everyday Human Videos which uses kinematic-only optimization retargeting for dexterous hands but doesn’t address the redundancy-consistency pathology that WARP identifies as a root cause of un-learnable data. Complements Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models on the offline-scaling side of the debate: Dyna-2 pays for retargeting noise with 1M+ hours of joint video+action training; WARP argues you can pay for it with a better solver instead. The whole-body-mobile-manipulation framing also puts it adjacent to Scaling Behavior Foundation Model for Humanoid Robots ScaleBFM but at the substrate level rather than the pretraining-recipe level.