Patent Landscape: 3D/4D Reconstruction × Robot Manipulation

Sources: Session notes, 2026-05-19 Raw: US Patent Landscape 3D Reconstruction Updated: 2026-08-12

Overview

A systematic US patent search covering “3D/4D generative methods × robot manipulation × safety assessment” confirms five white spaces with no corresponding granted or pending US patents as of May 2026. The methodology finding is equally important: robot ML patents rarely cite academic papers, making academic-citation-based prior art searches ineffective in this domain.

Methodology Note

Robot ML patents use broad functional claim language and minimal NPL (non-patent literature) citations — unlike biomedical or chemical patents. “Reverse citation search from academic papers” is largely ineffective here. Recommended search strategy: IPC classification codes + major assignee filtering.

Key IPC codes for this domain:

TechnologyPrimary IPC
Video / image generationG06T 13/xx
3D model / point cloudG06T 17/xx
Deep learning generative modelsG06N 3/045
Robot program controlB25J 9/16
Robot simulationB25J 9/1697
Robot safety devicesB25J 19/06
ML simulationG06F 30/27

Search pattern: IPC:B25J9/1697 AND IPC:G06N3/045 targets deep learning + robot simulation.

Existing Patents by Direction

Robot Safety (HRI Collision)

  • US12420422 — Situation-aware Safety Assessment of Robot-Human Activities: visual sensing → collision probability × severity → risk score → action control. Most direct granted case for “visual assessment of human-robot collision safety.”
  • US11872702B2 (Dexai, 2024) — Robot Interaction with Human Co-workers: 3D model geometric overlap + neural network collision prediction.

AI-Generated Safety Constraints

  • US20250042032A1 (pending, 2025) — LLM prompt + text environment description → safety constraints → control signal modification. Input is text, not visual image.
  • US20240253224A1 / US20240359319A1 (Sanctuary AI, 2024) — LLM-driven robot system, IPC safety device classification.

Simulation-Driven Control

  • US11938638B2 (Google/GDM, 2024) — Simulation Driven Robotic Control: 3D point cloud simulated environment → candidate action evaluation → real robot execution. Closest granted case to “pre-execution rollout” concept. Uses static point clouds, not generative.

Transformer + VLM → Action

  • US20240189994A1 (Google, 2024) — Real-world Robot Control Using Transformer NNs (RT series)
  • US20240078429A1 (Google, 2024) — Control Policies for Robotic Agents: learned predictive model + MPC

Diffusion Policy

  • US20250312914A1 (2025) — Transformer Diffusion for Robotic Task Learning
  • US20250353169A1 (2025) — Semi-supervised Learning of Robot Control Policies (NPL cites RT-1/RT-2, not Diffusion Policy original paper)

3D Reconstruction Foundations

  • US12145617B2 (NVIDIA) — 3D Surface Reconstruction with Point Cloud Densification Using AI

Confirmed White Spaces (as of May 2026)

  1. 3DGS / NeRF → robot manipulation planning — completely empty
  2. Video generation as pre-execution simulation — completely empty (V-Dreamer, Cosmos Policy, GPC exist only as academic papers)
  3. Point cloud generation for manipulation planning — nearly empty (US11938638B2 uses static point clouds, not generative)
  4. VLM visual assessment of robot action safety (image-input scoring) — empty; US20250042032A1 is closest but uses text input
  5. 3D/4D geometric world model × pre-execution safety scoring integrated pipeline — empty

Claim Strategy: VLM vs. Functional Language

Writing “VLM” directly in independent claims narrows protection scope — competitors can substitute equivalent architectures. Recommended layering:

Independent claim (broad):

“a multimodal machine learning model configured to process visual observations and generate structured evaluation outputs”

Dependent claim (narrow):

“wherein the multimodal machine learning model is a vision-language model (VLM)” “wherein the evaluation output comprises a numerical safety score and a natural language explanation”

Key Prior Art Distinctions

The most dangerous prior art for a claim combining “3D input → simulated trajectory → visual feature scoring”:

Patent3D InputSimulated TrajectoryVisual Feature Scoring
US11938638B2 (Google)✗ (state matching)
US20240025035A1 (Intrinsic)✗ (physical threshold)
US10766136 (Amazon)✓ (real execution only)
Target combination

Amazon US10766136: CNN analyzes real robot execution video → task success score → RL reward. Uses CNN (not visual+linguistic joint model); assesses real execution (not simulated pre-execution).

Distinction strategy for a visual pre-execution safety claim:

  1. Evaluation target: action feasibility/safety (not environment description)
  2. Output form: structured score + language explanation + downstream control signal
  3. Temporal position: pre-execution (not post-execution reward)
  4. Input modality: 3D point cloud or 4D video (not 2D image)

Open Questions

  • As 3DGS and NeRF-to-manipulation papers publish in 2025-2026, how quickly does the white space close? First-mover timing matters.
  • Does the VLM visual safety scoring white space extend to European (EPO) and Chinese (CNIPA) patent jurisdictions?
  • US20250042032A1 (text-input safety constraints) is pending — if granted, does its claim scope reach visual-input variants?

See Also