A manipulation foundation model built for industrial-grade generalization
Point-cloud transformer backbone. Force-torque feedback loop. Sub-100ms inference at the cell edge. Deployable on FANUC, KUKA, and ABB arms without reprogramming.
What the model does
Six core capabilities designed for high-mix, low-volume manufacturing where part variety is a constant.
Infers stable grasp candidates from point-cloud geometry alone. No CAD file required, no part-specific training cycle. Handles novel geometries at first encounter.
Runs on embedded GPU at cell edge. 80ms end-to-end from sensor input to joint command, within standard PLC scan cycle budgets.
Real-time force and torque sensing closes the manipulation loop. Detects slip, replans mid-grasp, and recovers failed picks without operator intervention.
Native SDK and REST interfaces for FANUC, KUKA, and ABB arms. Drop-in middleware layer; existing PLC and MES integrations continue unchanged.
Every production pick is a training signal. The model fine-tunes on cell-specific data, compounding catch-rate improvements over weeks without manual intervention.
Inference runs fully on-site. No part geometry, sensor data, or process information leaves the factory floor. ISO 27001 aligned deployment procedures.
How the model pipeline works
Four stages from raw sensor data to robot joint command. Each stage is independently observable and logged for production diagnostics.
Benchmark results on RMB-2K
Evaluated on the RLWRLD Manipulation Benchmark across 2,000 novel industrial parts. Compared against published baselines from CoRL 2025 proceedings.
| Model | Zero-shot grasp rate | Pose MAE (mm / deg) | Inference (ms) | F-T feedback |
|---|---|---|---|---|
| RLWRLD v1.2 | 94% | 1.8 / 2.1 | 80 | |
| GraspNet-1Billion | 78% | 4.2 / 5.6 | 210 | |
| Contact-GraspNet | 71% | 5.1 / 6.3 | 340 | |
| AnyGrasp | 81% | 3.6 / 4.2 | 185 |
Baselines re-evaluated on RMB-2K using authors' published checkpoints. All metrics reported on unseen part categories held out from pre-training. Details in our CoRL 2025 paper.
Add grasp intelligence in under a day
The RLWRLD SDK wraps the model inference API and controller translation layer. Three lines of code to connect your existing vision pipeline.
Native SDK for both Python 3.10+ and C++17. REST API for any other language. gRPC streaming for low-latency loops.
Tested with FANUC R-30iB+, KUKA KR C4, and ABB OmniCore controllers. Additional controller adapters available on request.
Works with Intel RealSense D455, Azure Kinect, and Photoneo MotionCam. Any ROS2 PointCloud2 topic accepted.
Optimized TensorRT runtime for Jetson AGX Orin (primary). x86 Docker container for simulation and development.
from rlwrld import ManipulationClient, GraspRequest
client = ManipulationClient(endpoint="http://edge-node:8080")
request = GraspRequest(
point_cloud=ros2_cloud_msg,
ft_reading=ft_sensor.current(),
task="pick_and_place",
)
grasp = client.predict(request)
print(grasp.pose, grasp.confidence)Deploy on your robot cell this quarter
We work directly with your engineering team through a structured 4-week onboarding program. Bring your part family; we bring the model.