SpatialDex: Does 3D Video Pretraining Help Bimanual Manipulation?.

Reconstruction-Augmented Visuomotor Pre-Training of 3D-Grounded Bimanual Dexterity from RGB-Only Human Videos • Xingyi He et al., including Kuan Fang • Cornell University • arXiv preprint • 2026.

Cold Open.

Picture a robot passing a flower from one hand to the other. The receiving hand has to arrive at the stem. Its fingers have to close at the right moment. The giving hand has to release only after the transfer works. A caption such as “pass the flower” leaves almost everything interesting unspecified.

Now imagine learning that coordination from ordinary human videos. Both hands move, the camera moves, and the objects exist at a physical scale that the pixels never directly state. Separating hand motion from camera motion is already a reconstruction problem.

SpatialDex builds its answer around a shared three-dimensional account of the hands and their surroundings. Passing a flower is among its real-robot evaluation tasks.

The central question is whether reconstructing that spatial context changes what a policy learns from video—and whether the difference becomes useful when two robot hands have to cooperate.

The Idea.

SpatialDex starts with a million human-video clips, drawn from Something-Something v2, EPIC-Kitchens-100, Ego4D, and Ego-Exo4D. Its vision-language backbone starts from pretrained PaliGemma weights. The new supervision couples reconstructed hand movement to reconstructed scene geometry.

That places the paper in a specific research lineage. VITRA already converts everyday human videos into instruction-conditioned hand-action examples and adapts the resulting model to robots. VIPA-VLA already uses human-hand information to calibrate reconstructed spatial annotations. These provide precedents for both action learning from video and hand-grounded scale recovery.

SpatialDex’s reconstruction tools are HaWoR for hands and VGGT-Omega for scenes and camera poses. Its important alignment decision comes afterward: use estimated hand size as an approximate ruler, and solve for one scale shared across the entire clip and both hands. Future hand trajectories and scene geometry are then expressed in the current camera’s coordinate frame.

Imagine that a kettle appears closer because the camera wearer leans forward. A useful action target should describe the reach toward the kettle, rather than accidentally encoding head movement as robot motion. Likewise, the reconstructed hand and kettle handle should inhabit the same units. The alignment pipeline makes those requirements explicit.

The policy then needs to combine spatial information with its existing visual knowledge. Any3D-VLA provides a useful architectural precedent: encode a point cloud in three dimensions, then associate its features with the corresponding image patches. That preserves the connection between what a patch depicts and where its contents are located.

SpatialDex follows this grid-aligned approach. Depth samples become three-dimensional points, a point-cloud encoder processes them, and their features are pooled back onto the visual token grid. A learned, dual-gated residual branch blends appearance and geometry, with its contribution introduced gradually during training. The original visual features remain available throughout.

Think of that as teaching an already competent visual model a new coordinate system. The visual representation can identify a handle; the geometric representation supplies spatial structure associated with that same region.

Geometry also becomes a prediction target. An auxiliary head learns future scene point sets alongside the action objectives. This rewards the shared representation for retaining information about how the surrounding scene evolves. The geometry predictor acts as a training-time teacher and is removed for deployment.

Finally, the action generator separates wrist motion from finger articulation. One diffusion-transformer expert predicts trajectories for both wrists. Another predicts both hands’ finger actions conditioned on that shared wrist plan. In a handover, finger closure therefore has access to where the giving and receiving hands intend to move.

Robot adaptation uses teleoperated demonstrations. At runtime, the policy receives RGB-D observations, camera calibration, language, and robot state. The video reconstruction machinery stays offline. “RGB-only” describes the human training videos; physical execution uses sensor depth through the same geometric input representation.

The Result.

The headline is 38.3 percentage points more real-robot task success than the strongest tested baseline, pi-zero-point-five. SpatialDex succeeds nearly twice as often under the paper’s adaptation protocol.

The hardware combines Wuji hands with Tianji arms. The six tasks span a flower handover, pouring, spraying a plant, unzipping a bag, sweeping, and packing a suitcase. Each method receives the same fifty teleoperated demonstrations per task. Evaluation uses ten trials per task and matched initializations; success means completing the task within its allowed execution time.

That task mix matters. Pressing a sprayer tests functional tool use. Unzipping couples two hands through a deformable object. A handover changes which hand supports the object. These are useful coordination probes because reaching the right neighborhood is only one stage of successful execution. The comparison includes scratch-trained policies, general pretrained VLAs, geometry-aware models, and human-video-pretrained policies.

Simulation broadens the evaluation across nine tasks, mixing single-arm and bimanual work on Allegro–Franka and Wuji–Tianji embodiments. Several tasks come from DexJoCo. That benchmark emphasizes functional interactions: its microwave task requires opening the door, placing food, closing the door, and starting the appliance. Success concerns a sequence of object-state changes.

SpatialDex’s largest simulated improvements occur on bimanual tasks. That pattern is consistent with its coordination hypothesis: the intended motion of both wrists gives finger control information about an evolving interaction, rather than just the current scene.

The most informative experiment is the matched pretraining comparison. The authors hold the videos, hand-motion labels, and pretraining budget fixed. Motion-only pretraining and full SpatialDex subsequently receive the same RGB-D robot architecture, demonstrations, objectives, and fine-tuning schedule. Adding reconstructed geometry as input and supervision during human-video pretraining improves success by 14.2 percentage points on the simulated ablation suite.

That control matters because both systems ultimately receive robot depth. The comparison tests the value of learning with geometry earlier, rather than simply rewarding a model for having depth available at execution time. It directly addresses the central pretraining question.

The component ablations reinforce this account. Shared clip-level alignment beats independent frame alignment. Explicit three-dimensional encoding beats treating depth as an image channel. Future-geometry supervision contributes beyond geometric inputs alone. Together, these results favor designing the representation and its training objectives as a coordinated system.

The Biggest Weakness.

The experiment isolating 3D human-video pretraining is simulated, while the headline hardware gain compares complete systems. The real-robot table contains SpatialDex and external baselines. The matched motion-only and no-human-pretraining controls appear in the simulation ablation table.

Those experiments answer different questions. The hardware comparison asks which adapted package works best. The matched comparison asks what reconstructed geometry contributes during pretraining. Combining them into “3D video grounding nearly doubles real-world success” assigns a hardware effect to an ingredient that was not isolated on hardware.

SpatialDex changes several things together: geometric representation, feature fusion, future-scene supervision, wrist–finger action generation, and human-video pretraining. These components form the integrated system tested against the external baselines.

For a lab deciding where to spend engineering effort, the distinction matters. Does reconstruction-based video pretraining account for most of the physical improvement? Or does a substantial share come from the architecture learning effectively from the robot demonstrations? The published hardware comparison cannot separate those explanations.

The simulation control supplies substantive evidence for geometry-rich pretraining. It does not establish the same effect size on physical hands. Sensor observations and contact-rich execution could change the relative value of the learned prior. This is a question about transferring a causal result between evaluation settings, not an allegation that the robot successes are invalid.

The decisive follow-up is straightforward: run the existing matched pretraining variants on the physical platform. Keep the RGB-D inputs, robot demonstrations, action heads, adaptation schedule, and evaluation initializations fixed. Compare full geometry pretraining with motion-only pretraining.

That would directly measure the hardware return on reconstructing the videos. The paper establishes a strong integrated system and a persuasive simulated mechanism. The missing link is a controlled hardware test of that mechanism.

Verdict.

SpatialDex deserves attention from teams building bimanual policies with multifinger hands and RGB-D observations. Its practical contribution is a recipe for making human-video supervision use the same spatial relationships as the robot’s observations and actions. Scene reconstruction becomes part of learning the motor prior.

The useful shift is in the unit of training data. A wrist trajectory is a trace through space. Pair that trajectory with the surrounding scene, and training can connect the motion to a handle, a receiving hand, or a container. My engineering takeaway is to preserve those relationships as first-class supervision.

For representation-learning researchers, the compelling design principle is the pairing of input and objective. Supply spatial information in a form the backbone can associate with its visual tokens, then reward the representation for preserving information about future spatial structure. SpatialDex’s ablations support that combination.

There is useful infrastructure around this research direction. VITRA’s public repository supplies human-video annotations, pretrained models, and a custom-robot fine-tuning path. DexJoCo provides replayable demonstrations and executable manipulation evaluations. Any3D-VLA supplies an architectural precedent for aligning point-cloud features with visual patches. These are concrete starting points for investigating geometry-rich video learning.

The bimanual setting also sharpens the scientific question. In a handover, one hand’s target is partly determined by the other hand’s behavior. The destination is itself controlled. A shared wrist plan is therefore a useful organizing variable for reasoning about the interaction, and SpatialDex explicitly gives that plan to the finger expert.

That suggests a productive research agenda: organize pretraining around relationships that remain meaningful across bodies—approaching, supporting, receiving, releasing—and give those relationships a consistent geometric context.

The takeaway: learn hand motion together with the three-dimensional scene that makes that motion useful.

Outro.

Return to the flower. The useful knowledge is more than the shape of a grasp. It is how that grasp belongs to a moving relationship between two hands and an object.

SpatialDex makes that relationship available to training through aligned scene geometry, future-scene supervision, and coordinated action prediction.

The next development to watch is whether scene-and-action reconstruction becomes a reusable pretraining format alongside images, instructions, and robot trajectories.

This has been Embodied AI 101. For a robot passing a flower, knowing where both hands are going is part of knowing when to let go.