lab NVIDIA Research · 1h ago · research

ROSA improves factory productivity by up to 12.06x via shared GPU-pool serving

from English

Researchers propose ROSA, a robotics foundation model serving system designed for robot factories that moves beyond single-robot, edge-computing assumptions. The system utilizes shared GPU-pool serving to allow fleets of robots to access server-class GPUs over the network.

ROSA adopts shared GPU-pool serving to improve inference performance, battery duration, and GPU utilization across a fleet.
It provides a robotics-aware programming abstraction supporting multi-model pipelines, per-task performance requirements, and failure handling.
The system uses factory-objective-driven scheduling to maximize SLO-qualified factory productivity rather than minimizing individual request latency.
Implemented on Ray Serve with vLLM, PyTorch, and JAX backends, ROSA improves factory productivity by up to 12.06x over conventional dedicated serving systems.

ROSA addresses the limitations of existing systems by optimizing for collective factory outcomes instead of isolated robot latency.

Importance 1/3 Trust 3/3 NVIDIA Research Research paper Robotics

Read original

arxiv arXiv cs.AI · 24d ago · 1 view

PAC-ACT post-training framework improves ACT policies for precision contact manipulation

This paper introduces PAC-ACT, a reinforcement-learning post-training framework designed to enhance pretrained Action Chunking Transformer (ACT) policies for industrial precision-contact tasks. The method reformulates policy optimization at the chunk level and utilizes an ACT-transferred actor-critic architecture with a hybrid behavior-prior constraint.

arxiv arXiv cs.AI · 26d ago · 4 views

Track2Map: online deformable SLAM for robotic surgery

Researchers propose Track2Map, an online 3D Gaussian Splatting pipeline that jointly optimizes camera trajectory and 3D scene representation directly from surgical video. This approach enables robust reconstruction in robot-assisted minimally invasive surgery even when accurate camera trajectory priors are missing or noisy.

arxiv arXiv cs.LG · 30d ago · 3 views

LIME learns intent-aware camera motion from egocentric video

The authors propose LIME, a vision-language model that generates language-conditioned camera motion by predicting relative target poses from RGB observations and natural-language intents.

arxiv arXiv cs.AI · 30d ago · 3 views

WorldSample uses closed-loop RL with world modelling to improve real-robot training

Researchers propose WorldSample, a physically grounded data augmentation framework that closes a real-synthetic loop between physical rollouts, world-model generation, and policy improvement for reinforcement learning on real robots. The system generates high-fidelity synthetic transitions through a post-trained world model and employs Policy-Paced Learning to regulate training via sample selection, balancing useful augmentation against value overestimation.

arxiv arXiv cs.CL · 33d ago · 1 view

PhysMani: Physics-principled 3D World Model for Dynamic Object Manipulation

Researchers propose PhysMani, a framework that couples a physics-principled 3D Gaussian world model with a future-aware action policy model to address challenges in manipulating fast-moving targets in unstructured 3D environments.