arXiv 2026 ยท Robotics ยท VLA

3DVLA: Enhancing Vision-Language-Action Models via 3D Spatial and Instance Understanding

VLARobotic manipulation 3D perceptionMulti-view Self-supervised

TL;DR

3DVLA is a plug-and-play 3D add-on for pretrained 2D Vision-Language-Action models. It bolts on (1) multi-view spatial fusion with 3D positional encoding, (2) object-centric 3D instance probes, (3) a masked self-supervised predictor that hallucinates occluded geometry, and (4) an uncertainty-gated geometry aggregator. With only ~0.086B extra parameters, it reaches 86.0% average on LIBERO-Plus (vs. 84.2% for ฯ€โ‚€.โ‚…) and lifts ฯ€โ‚€ from 46.4% โ†’ 54.5% (Easy) and 16.3% โ†’ 23.2% (Hard) on RoboTwin 2.0.

Problem & Motivation

Vision-Language-Action (VLA) models like OpenVLA, ฯ€โ‚€, and ฯ€โ‚€.โ‚… inherit huge 2D vision-language pre-training, but their understanding of the 3D scene is thin. The paper identifies three coupled failures:

  1. Weak 3D spatial extraction. Existing VLAs encode single views or loosely fuse multiple views without ever reconstructing a metric 3D frame, so they cannot reason about precise 6-DoF poses.
  2. Insufficient 3D instance understanding. Objects are treated as flat image regions; the model has no volumetric extent, no 3D boundary, and cannot reliably tell which pixels belong to which object in 3D.
  3. Fragile reasoning under occlusion. On LIBERO-Plus, simply nudging the camera so the target is partially occluded crashes SOTA VLAs' success rates โ€” a purely-2D model cannot infer the missing geometry.

Off-the-shelf 3D perception backbones (ImVoxelNet, R4Det, etc.) could fix this, but plugging them into a VLA forces a redesign that discards 2D VLM pre-training and demands dense 3D instance annotations that don't exist at robotic-data scale. 3DVLA aims to retain VLM priors, require no extra manual labels, and still inject robust 3D reasoning.

Key Concepts

Multi-View Spatial Fusion
Flattens \(V\) view feature maps into one unordered token sequence, adds a learned absolute-3D positional embedding from unprojected coordinate grids, and runs a Transformer encoder. Output: a globally consistent 3D memory \(\mathcal{M}\).
3D State Probes
Learnable queries \((\mathbf{c}_j,\mathbf{p}_j)\): a semantic vector plus an explicit 3D anchor point. They are projected into each view to do multi-scale deformable sampling, then refined with occlusion-aware gating and a tanh-bounded coordinate update.
Global 3D Hungarian Matching
Pseudo-GT 2D boxes/masks come from a frozen open-world detector (VL-SAM v2); their 3D centroids are obtained by unprojecting with depth. A single Hungarian assignment in 3D eliminates the per-view ID conflicts that plague 2D matching.
Coordinate-Driven 3D Predictor
An I-JEPA-style masked self-supervised branch โ€” but instead of discarding the predictor after pre-training, 3DVLA keeps it and reuses it at inference to synthesize visual tokens for occluded 3D locations.
Continuous 3D RoPE
Rotary positional encoding applied to queries/keys whose phase angle \(\theta_d=p\cdot 10000^{-2d/D_{axis}}\) is computed continuously per axis \(p\in\{x,y,z\}\), giving the attention a relative-distance prior in 3D.
Uncertainty-Guided Routing
A gate that absorbs synthesized 3D context into the instance representation only when the 2D classifier is unsure (\(u_i=1-\sigma(l_i)\) is high), zero-initialized with bias \(-3.0\) so training starts as an identity mapping.

The Method

3DVLA sits between the frozen 2D vision encoder and the action expert of an existing VLA. The four modules form a pipeline that lifts 2D tokens into 3D, extracts entities, repairs occluded geometry, and feeds spatially-grounded instance tokens to the policy.

Multi-view 2D features Spatial Fusion + 3D pos. emb. โ†’ memory ๐“œ 3D Instance Module state probes (๐œโฑผ,๐ฉโฑผ) global 3D Hungarian 3D SSL Predictor masked โ†’ completion H-Attn w/ RoPE-3D Geom. Aggreg. EE-relative PE uncertainty gate to action expert
The four components of 3DVLA: multi-view spatial fusion produces a globally consistent 3D memory; the instance module decodes object-centric tokens; the masked SSL predictor synthesizes occluded tokens; the aggregator fuses local 3D context into each instance token under an uncertainty-driven gate.

The pieces

1. Multi-View Spatial Fusion

Flatten \(\mathcal{X}_{fmap}\in\mathbb{R}^{V\times C\times H\times W}\) into \(N=V\!\cdot\!H\!\cdot\!W\) tokens; add a 2-layer MLP+LayerNorm 3D position embedding derived from unprojected world coordinates; pass through a Transformer encoder to enforce cross-view consistency.

2. Object-Centric 3D Instance Module

Instantiate \(N_q\) state probes in continuous 3D; project each \(\mathbf{p}_j\) into every view with depth-clamped projection; multi-scale deformable sampling; tanh-bounded coord refinement; global Hungarian matching in 3D against pseudo-GT from a frozen VL-SAM detector.

3. Coordinate-Driven SSL Predictor

Asymmetric multi-view masking forces cross-view reasoning. A Hybrid Attention block uses RoPE-3D on Q/K plus Fourier 3D pos-enc + view embedding added to the values. The predictor is kept after pre-training to complete tokens at occluded coordinates.

4. Spatially-Conditioned Geometry Aggregation

End-effector-relative positional encoding gives the action expert offsets in gripper-centric coordinates. An uncertainty-guided gate (\(u_i=1-\sigma(l_i)\)) only mixes 3D geometry into instance tokens when the 2D classifier is unsure, preserving confident features.

Math Walkthrough

Lifting 2D tokens to 3D

Per-view tokens are flattened to \(\mathcal{M}_{flat}\in\mathbb{R}^{N\times C}\) and given an explicit 3D position embedding:

$$\mathbf{E}_{3D}=\operatorname{LayerNorm}\!\Big(\mathbf{W}_2\,\operatorname{GELU}(\mathbf{W}_1\mathbf{P}_{flat}+\mathbf{b}_1)+\mathbf{b}_2\Big)$$

where \(\mathbf{P}_{flat}\) holds the unprojected world coordinates of every token. Continuous 3D RoPE is then applied on attention queries and keys with per-axis phases \(\theta_d=p\cdot 10000^{-2d/D_{axis}}\) for \(p\in\{x,y,z\}\).

Instance probes & view projection

For each probe \((\mathbf{c}_j,\mathbf{p}_j)\) and view \(v\) with intrinsics \(\mathbf{K}_v\) and extrinsics \([\mathbf{R}_v\,|\,\mathbf{t}_v]\):

$$\mathbf{u}_{j,v}=\Phi_{norm}\!\left(\frac{\mathbf{K}_v(\mathbf{R}_v\mathbf{p}_j+\mathbf{t}_v)}{\max(Z_c,\epsilon)}\right)\in[0,1]^2,\quad\epsilon=0.01$$

The depth clamp prevents numerical blow-up near the camera plane. Multi-scale deformable sampling around \(\mathbf{u}_{j,v}\) gives the view-specific feature:

$$\mathbf{h}_{j,v}=\sum_{m=1}^{M}\mathbf{W}_m\sum_{k=1}^{K}\sum_{s=1}^{S}A_{j,v,m,k,s}\cdot\mathbf{W}'\mathcal{M}_{ms}^{(s)}\!\left(\mathbf{u}_{j,v}+\Delta\mathbf{u}_{j,v,m,k,s}\right)$$

Bounded coordinate refinement

An occlusion-aware gate \(g_{j,v}=\sigma\!\big(\operatorname{MLP}(\mathbf{h}_{j,v})/\tau_{view}\big)\) suppresses bad views. The 3D anchor is updated with a hard physical bound:

$$\mathbf{p}_j^{(l+1)}=\mathbf{p}_j^{(l)}+\alpha\cdot\tanh\!\Big(\operatorname{MLP}_{coord}(\mathbf{c}_j^{(l+1)})\Big)$$

The \(\tanh\) keeps per-layer displacement bounded by \(\alpha\), guaranteeing stable convergence.

Global 3D matching & loss

$$\mathcal{C}_{global}(j,k)=\frac{1}{\max(|\mathcal{V}_k|,1)}\sum_{v\in\mathcal{V}_k}\mathcal{C}_v(j,k)$$
$$\mathcal{L}=\lambda_{cls}\mathcal{L}_{cls}+\lambda_{box}\mathcal{L}^{2D}_{L1}+\lambda_{giou}\mathcal{L}_{GIoU}+\lambda_{mask}\mathcal{L}^{mask}_{CE}+\lambda_{dice}\mathcal{L}^{mask}_{DICE}+\lambda_{3d}\mathcal{L}^{3D}_{L1}$$

Hybrid Attention for occlusion completion

Context \(\tilde{\mathcal{X}}_{vis}=\mathcal{X}_{vis}+\Phi_{pos}(\mathbf{P}_{vis})+\mathbf{E}_{view}(\mathbf{V}_{vis})\) and target \(\tilde{\mathcal{Z}}_{tgt}=\mathcal{Z}_{tgt}+\Phi_{pos}(\mathbf{P}_{tgt})+\mathbf{E}_{view}(\mathbf{V}_{tgt})\). RoPE-3D is applied only to Q and K:

$$\mathbf{Q}=\operatorname{RoPE}_{3D}(\mathcal{X}_q\mathbf{W}_Q,\mathbf{P}_q),\quad\mathbf{K}=\operatorname{RoPE}_{3D}(\mathcal{X}_{kv}\mathbf{W}_K,\mathbf{P}_{kv})$$
$$\operatorname{H\text{-}Attn}(\mathcal{X}_q,\mathcal{X}_{kv})=\operatorname{softmax}\!\Big(\tfrac{\mathbf{Q}\mathbf{K}^\top}{\sqrt{d_k}}\Big)(\mathcal{X}_{kv}\mathbf{W}_V)$$

Uncertainty-guided routing

$$\mathbf{g}_i=\sigma\!\Big(\tau\cdot\operatorname{MLP}_{gate}\!\big([\mathbf{c}_i\,\|\,\mathbf{h}_i\,\|\,u_i]\big)\Big),\quad u_i=1-\sigma(l_i)$$

The gate is zero-initialized with bias \(-3.0\) so \(\mathbf{g}_i\approx\mathbf{0}\) at the start of training, leaving the pretrained 2D features untouched. The final instance representation is \(\hat{\mathbf{c}}_i=\mathbf{c}_i+\mathbf{g}_i\odot\operatorname{LayerNorm}(\mathbf{h}_i)\).

Results

LIBERO-Plus (zero-shot transfer from LIBERO)

Success rate (%) across the seven perturbation dimensions plus average. Bigger is better.

MethodCameraRobotLang.LightBg.NoiseLayoutAvg.
OpenVLA0.83.523.08.134.815.228.515.6
NORA2.237.065.145.758.612.862.139.0
UniVLA1.846.269.669.081.021.231.942.9
ฯ€โ‚€-Fast65.121.661.073.273.274.468.861.6
RIPT-VLA55.231.277.688.491.673.574.268.4
OpenVLA-OFT56.431.979.588.793.375.874.269.6
ฯ€โ‚€61.040.863.589.384.180.176.469.4
ฯ€โ‚€.โ‚…71.277.189.994.794.084.284.384.2
ฯ€โ‚€.โ‚… + 3DVLA75.677.488.697.497.785.386.686.0

RoboTwin 2.0 (physics-rich, contact-heavy)

MethodEasyHard
DP28.0%0.6%
ACT29.7%1.7%
RDT34.5%13.7%
DP355.2%5.0%
ฯ€โ‚€46.4%16.3%
ฯ€โ‚€ + 3DVLA54.5%23.2%
X-VLA70.0%39.0%
X-VLA + 3DVLA72.6%42.1%

Ablation on RoboTwin 2.0 (built on ฯ€โ‚€)

ModelInst.Pred.Rout.EasyHard
ฯ€โ‚€ (baseline)โ€”โ€”โ€”46.4%16.3%
Aโœ“โ€”โ€”50.1%18.9%
Bโœ“โœ“โ€”53.4%21.8%
C (ours)โœ“โœ“โœ“54.5%23.2%

Micro-ablations on LIBERO-Plus: removing the voxel-free spatial fusion drops average from 85.0% โ†’ 84.6%; swapping global 3D Hungarian for 2D matching โ†’ 84.5%; removing end-effector relative PE โ†’ 84.8%. Masking ratio: 25% gives 85.5% (the network short-circuits by interpolating from neighbors); 50% gives 85.7% โ€” proving asymmetric heavy masking is what forces real cross-view 3D reasoning. 3DVLA adds only ~0.086B parameters.

Intuition & Analogies

Why a separate 3D module beats "just train on more views". A 2D VLA sees several pictures of the same scene but has no enforced rule that says "this red pixel in view A is the same physical point as that red pixel in view B". The Spatial Fusion module is the rule: every token is tagged with an explicit world-coordinate embedding before it ever talks to the others, so cross-view consistency falls out of the attention math instead of having to be discovered from data.

The predictor as an "imagination organ". Masked self-supervised methods like I-JEPA throw away the predictor head once pre-training is done โ€” it was just a training-time scaffold. 3DVLA's insight is that, once the predictor is conditioned on 3D coordinates, it doubles as a runtime tool for hallucinating the parts of objects you can't see right now. The Camera-perturbation gains (71.2 โ†’ 75.6, and 71.4 โ†’ 73.6 in the ablation) are exactly the slice this is built for.

Uncertainty gating as a humility check. If the 2D detector is confident, the 3D context is a distraction and could corrupt good features. The zero-initialized routing gate (\(\mathbf{g}_i\approx\mathbf{0}\) at start, bias \(-3.0\)) means the 3D branch has to earn the right to write into the instance token โ€” and it does so only when the 2D evidence is ambiguous.

Limitations & Open Questions

  • Depends on calibrated depth and camera matrices. The instance probe projection uses \(\mathbf{K}_v\) and \([\mathbf{R}_v\,|\,\mathbf{t}_v]\) explicitly; deploying on a robot without good calibration or depth would weaken the geometry signal.
  • Pseudo-GT from a frozen detector. 3D centroids are derived from VL-SAM v2 masks unprojected with depth. Errors in the open-world detector propagate into the matching cost.
  • Simulation benchmarks only. Results are reported on LIBERO-Plus and RoboTwin 2.0 (simulation); the paper does not report real-robot evaluations.
  • Modest absolute gains in Language and Robot splits. On LIBERO-Plus the Language split actually dips slightly (89.9 โ†’ 88.6), suggesting the 3D context is not always helpful when the bottleneck is linguistic.
  • RoboTwin Hard remains low in absolute terms. Even with 3DVLA, the best ฯ€โ‚€-based number is 23.2%; the X-VLA-based 42.1% leaves a wide gap to robust manipulation.

Takeaways

  • 3D reasoning can be bolted onto a 2D VLA without retraining the VLM. Plug-and-play gains on ฯ€โ‚€, ฯ€โ‚€.โ‚…, and X-VLA.
  • Enforce multi-view consistency at the token level โ€” adding a learned 3D position embedding before fusion outperforms decoder-side tricks.
  • State probes carrying explicit 3D anchors + global 3D Hungarian matching avoid the cross-view ID conflicts that 2D matching reintroduces.
  • Don't throw the masked-SSL predictor away. Keep it and reuse it at inference to fill in occluded geometry โ€” this is the single biggest win on the Camera-perturbation split.
  • Gate 3D context by 2D uncertainty. A zero-initialized routing gate protects the pretrained features and only adds 3D information when the policy actually needs it.