When a person teleoperates a humanoid robot with their own body, their legs do not match the robot's legs. The strides differ, the limb proportions differ, the hips sit at different heights. Any mapping that converts the operator's pose into a robot pose has to absorb this mismatch somewhere. Current mappings absorb part of it at the planted foot, so the robot's foot slides across the floor while the human's foot stands still. This is foot sliding. In a teleoperated robot it causes slips and makes the robot track its reference motion badly. We propose to absorb the mismatch only where the human body is moving. The robot's body is placed by a continuously updated smooth, invertible warp of 3D space, a diffeomorphism. The warp is allowed to change only in regions where the operator's body is moving. A foot the human holds still then maps to a foot the robot holds still, by construction. This needs no contact detector and no binary "foot is down" switch, and the warp can never tear or fold.
This matters because the best whole-body teleoperation cockpits have stopped short of the legs. HOMIE (HOMIE) controls the arms isomorphically, meaning joint for joint through a leader device that copies the robot's joints, at about 300 Hz. Locomotion goes to foot pedals that drive a reinforcement-learning walking policy, so the operator cannot choose where a foot lands. Retargeting methods that do suppress sliding have different problems. Some run offline, like DynaRetarget (DynaRetarget) at about one minute of compute per second of motion. Others need offline contact labels and a one-second pause before each contact change (Multi-contact motion retargeting). We want an online mapping for the non-isomorphic part of the body, the legs and pelvis, that keeps what isomorphic arm control is valued for. That means continuity, no clutching or re-indexing, and no sudden jumps, plus a structural guarantee against sliding.
Core Claim
Let $x_i(t)$ be the world position of operator keypoint $i$ (feet, ankles, knees, pelvis, hands) at time $t$. Let $\Phi_t$ be the current warp from operator space to robot space, so the robot keypoint is $y_i = \Phi_t(x_i)$. Differentiating gives
$$\dot y_i = D\Phi_t(x_i)\,\dot x_i + u_t(y_i),$$
where $D\Phi_t$ is the 3×3 Jacobian of the warp and $u_t$ is the velocity field that is currently reshaping the warp. The first term is the operator's motion carried over to the robot. The second term is the correction that fits the motion to the robot's body. Foot sliding is exactly the case where $\dot x_i = 0$ but $u_t(y_i) \neq 0$.
Per-frame inverse kinematics re-solves the robot pose from scratch at every frame, and pelvis-relative scaling rescales limb positions around a moving hip. Both add a nonzero second term at the stance foot, which is why they slide. We instead multiply the correction by a smooth stillness mask, $u_t(y) = m_t(y)\,\tilde u_t(y)$. The mask $m_t \in [0,1]$ is spread through space by a kernel from the operator's keypoint speeds, $m \approx 1 - \exp(-\|\dot x_i\|^2/\sigma^2)$, so it goes to zero near any body point the operator holds still. Here $\tilde u_t$ is the unmasked correction field: a kernel interpolation of the residual between the warped keypoints and a robot-feasible pose from a warm-started differential inverse kinematics step.
The warp is advanced by composing small steps, $\Phi_{t+\Delta t} = (\mathrm{Id} + \Delta t\, u_t)\circ\Phi_t$. Each step is a guaranteed diffeomorphism whenever
$$\Delta t \cdot \mathrm{Lip}(u_t) < 1,$$
where $\mathrm{Lip}$ is the field's Lipschitz constant, its maximum stretch rate. This is the standard condition under which a perturbation of the identity stays invertible. We enforce it by clipping kernel amplitudes. The determinant of $D\Phi_t$ therefore stays positive forever, so the robot's body region can never fold onto itself or swap the ordering of nearby limbs.
The claim has three parts. First, stance-phase sliding of the robot reference is bounded by $m\cdot\|\tilde u\|$, and goes to zero as the operator's foot stops. Second, because the mask is continuous rather than binary, contact transitions produce no snapping. Third, the whole update is cheap enough for a teleoperation loop. We only need $\Phi_t$ and its Jacobian at a few dozen keypoints. These evolve by $\dot y_i$ above and by $\tfrac{d}{dt}D\Phi_t = Du_t\, D\Phi_t$, so no volumetric field is ever stored.
Relation to Prior Work
Three works supply the method. VoxelMorph (Unsupervised learning of probabilistic diffeomorphic registration) showed that flow-based warps reduce folding from thousands of voxels to about 0.2 on brain MRI at sub-second speed. We reuse the core lesson that invertibility should come from how the warp is built rather than from a penalty. We replace its stationary, image-sized field with a time-varying, keypoint-sized one. Landmark-constrained transport (Coupled Optimal Transport with Landmark Constraints) supplies the idea of pinning selected points during a deformation. In our case the pins are not fixed in advance; the operator's own stillness sets them moment by moment. The zero-velocity contact loss of Xia et al. is the binary ancestor of our mask. We keep its target, stationary stance feet, but drop the predicted contact label, which the file notes snaps on irregular or compliant ground.
Four works motivate the problem by leaving it open. Motion Mappings for Continuous Bilateral Teleoperation is the closest prior work. It learns a diffeomorphic map between two Panda arm workspaces with a normalizing flow, an invertible neural network trained to map one space onto another. It halved valve-turning time, from 195 s to 96 s. That map is time-invariant and trained offline for a static scene. A fixed map cannot follow an operator who walks without bound, because any fixed scaling of the world eventually pulls the robot back toward a center. Making the map pelvis-relative brings back the sliding term $\partial\Phi/\partial t$. Our warp evolves online for exactly this reason. HOMIE motivates the regime: it is the evidence that isomorphic control pays off, and also that the legs remain uncontrolled. IKMR (Implicit Kinodynamic Motion Retargeting) is fast at 5000 FPS but processes batch windows that leave seams between them. AdaMorph (AdaMorph) handles 12 robot embodiments but has no invertibility guarantee and runs offline. Neither is a causal streaming mapping.
Two works constrain the design. GMR (Retargeting Matters) shows that the reference matters a great deal: with flawed references, PHC's tracking success drops to 0% on dance and kung-fu motions. So we must report how well the robot tracks our references, not only how they look. GeoRT (Geometric Retargeting) shows that a geometrically regularized network mapping human hands to robot hands reaches 1000 Hz without formal guarantees. This sets a high bar for speed: our mapping has to be fast as well as guaranteed.
Research Program
We first build the streaming mapper. It has four parts: per-keypoint Lagrangian tracking of positions and Jacobians, the kernel correction field, the stillness mask, and the Lipschitz clipping. A warm-started differential inverse kinematics solver in the style of GMR's Mink stage then turns warped keypoints into joint targets for a Unitree G1. Because stance targets are stationary, the warm start keeps stance joints fixed.
Second, we evaluate in simulation, using motion-capture streams as a stand-in for a live operator. Each sequence is fed strictly frame by frame with no lookahead, and the output goes to a motion-tracking policy trained in the style of GMR's BeyondMimic evaluation.
Third, we build a cockpit that keeps HOMIE's isomorphic arm leader and replaces the pedals with lower-body motion capture feeding our mapper. We test it on foothold-placement and stepping-over tasks that pedals cannot express.
Fourth, we analyze the main tradeoff the formulation makes explicit. The Lipschitz bound limits how fast the warp can correct the motion between a pinned foot and a nearby moving pelvis. A tight bound means a guaranteed warp but slow correction of the body's shape. We will measure how tracking error and residual sliding trade off as $\sigma$ and the kernel width vary.
Baselines
The naive reference is direct scaled retargeting. Robot keypoints are operator keypoints scaled about the pelvis by the ratio of leg lengths, then passed through the same inverse kinematics solver. This is what the sliding term predicts will fail. It shows how much sliding there is to remove in the first place.
The strongest online competitor is GMR's pipeline (Retargeting Matters), restricted to run causally per frame. It keeps non-uniform segment scaling and differential inverse kinematics, plus foot equality constraints switched on by a speed threshold. We give it the same keypoints, solver, robot model, tracking policy, and a matched tuning budget for its threshold and weights as for our $\sigma$ and kernel width. GMR is the strongest competitor because it is the published method that most directly removes sliding and joint spikes on the G1. The file gives no numeric tracking error for GMR: "no reported result found on Lacuna for GMR $E_{g\text{-}mpbpe}$/$E_{mpjpe}$". The only reported comparative number is that PHC drops to 0% success on complex motions. Beating GMR means three things: lower stance-phase foot speed, no transition snapping as measured by peak foot acceleration at contact changes, and tracking success that is equal or better, all at teleoperation rates.
As an offline upper bound we cite DynaRetarget on OmniRetarget. It reports 74.6% success against SPIDER's 37.9%, and smoothness of 1.7 against 3.4, at about one minute of compute per second of motion. We do not expect to match its success rate online. Coming close at more than a thousand times less compute would be the meaningful result.
For the cockpit we compare against HOMIE-style pedal locomotion on identical tasks.
The key ablation is a per-limb velocity integrator. Each foot target integrates a scaled copy of the operator's foot velocity, so stillness is preserved. There is no shared spatial field and no invertibility guarantee. Comparing against it separates the effect of "preserve zero velocity" from the effect of "preserve it inside one coupled, non-folding warp". We also ablate the mask by replacing it with a hard speed threshold, and ablate the Lipschitz clipping by removing it.
Why not just…?
Why not just run GMR-style inverse kinematics online with a foot-speed contact threshold? This is the real threat, and it may come close on sliding alone. The file documents what goes wrong with binary contacts. They snap on irregular ground (Xia et al.), and online contact switching without force sensing destabilized the robot in Multi-contact motion retargeting. The threshold-mask ablation and the matched-budget GMR baseline settle the question directly. If thresholded inverse kinematics matches us on sliding, snapping, and tracking, then the diffeomorphic machinery is not earning its keep, and we will say so.
Why not just keep HOMIE's pedals and let the walking policy handle the feet? For flat-ground navigation, pedals are probably enough, and HOMIE reports 73.3% and 80.0% success on its learned tasks. Pedals specify velocity, not footholds, though. Our cockpit tasks require stepping onto marked targets and over obstacles, and pedals cannot express those commands at all. The comparison is on foothold error and task success there, not on open-floor walking.
Why not just learn a fixed invertible map, as Motion Mappings does for arms? A fixed world-frame map cannot follow unbounded walking. A body-relative fixed map brings back the $\partial\Phi/\partial t$ sliding term every time the pelvis moves. We include a pelvis-relative normalizing flow fitted to the same data as a baseline, so this argument is tested rather than assumed.
Why not let the tracking policy absorb the sliding, as LAGOON does with its physics-based tracker (LAGOON)? The tracking policy may hide sliding in the reference. GMR's 0% PHC result says it cannot always do so. We report sliding both in the reference and in the motion the robot actually executes, together with tracking success, so that any masking effect is visible.
Proposed Experiments
The simulation study streams the 21 LaFAN1 sequences used by GMR, a locomotion subset of AMASS, and the 285 OmniRetarget loco-manipulation motions used by DynaRetarget into each mapper, causally. Ground-truth contact labels from the human data are used only for scoring. The metrics are mean horizontal stance-foot speed and peak foot acceleration at contact transitions, both for the reference and for the executed motion. We also measure ground penetration, joint-velocity spikes, the minimum Jacobian determinant over the episode, per-frame latency, and tracking success with global and joint errors under the shared tracking policy. On OmniRetarget we additionally report DynaRetarget's success and smoothness definitions, so the results are comparable with its 74.6% and 1.7.
The decisive checks are three. First, as the mask width $\sigma$ shrinks, stance sliding should fall toward zero while tracking error rises only moderately. Second, the shared-field method should beat the per-limb integrator on how well pelvis and feet stay consistent and on tracking success. Third, the soft mask should beat the hard threshold on transition snapping without losing ground on sliding.
The hardware study runs the G1 with the arm leader and leg capture, at a control loop of at least 100 Hz. Novice operators perform foothold-placement, stepping-over, and squat-pick tasks against the pedal cockpit. We measure foothold error, completion time, falls, and practice attempts needed to reach a plateau. The learning-curve measure is modeled on HOMIE's report that novices reach expert speed within five attempts.
References
- HOMIE: Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit
- DynaRetarget: Dynamically-Feasible Retargeting using Sampling-Based Trajectory Optimization
- Multi-contact motion retargeting from human to humanoid robot
- Unsupervised learning of probabilistic diffeomorphic registration for images and surfaces
- Coupled Optimal Transport with Landmark Constraints
- Combining Recurrent Neural Networks and Adversarial Training for Human Motion Synthesis and Control
- Motion Mappings for Continuous Bilateral Teleoperation
- Implicit Kinodynamic Motion Retargeting for Human-to-humanoid Imitation Learning
- AdaMorph: Unified Motion Retargeting via Embodiment-Aware Adaptive Transformers
- Retargeting Matters: General Motion Retargeting for Humanoid Motion Tracking
- Geometric Retargeting: A Principled, Ultrafast Neural Hand Retargeting Algorithm
- LAGOON: Language-Guided Motion Control