AI Daybook
한국어
AI Daybook
Today's issue
PaperPhysical AI

RPG framework improves robot manipulation success to 95% through offline practice

Original titleReconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents

Outlet
arXiv
Published
Importance
4/5
Trending
6

Summary

The authors present RPG, a framework that identifies manipulation capabilities in offline datasets and constructs related practice tasks in simulation. On 22 held-out manipulation tasks, RPG improves success from 28.6% after the first practice round to 95.0% after 15 rounds, outperforming baselines including ASPIRE and CaP-Agent0, and all 30 physical trials succeed after hardware adaptation.

Also reported by 4 outlets

Original excerpt

Building reliable robot capabilities across diverse tasks requires substantial human effort to develop and maintain skills, design rewards, and integrate perception with control. We present Reconstruct, Practice, Go Real (RPG), a framework for autonomous improvement of robot execution systems without updating model weights. RPG identifies manipulation capabilities in an offline dataset and constructs related practice tasks in simulation. During practice, RPG uses execution feedback, privileged simulator state, and available dataset videos to diagnose failures. It develops new reusable symbolic skills, refines existing skills, and revises the system prompt based on these diagnoses. Cross-task evaluation tests individual candidate changes and merged revisions before they are retained for reuse. At test time, a multimodal LLM uses the resulting system prompt and skill library to coordinate perception and robot control. On held-out initializations of 22 manipulation tasks, RPG improves task success from 28.6% after the first practice round to 95.0% after 15 rounds, outperforming all evaluated baselines, including ASPIRE (75.5%) and CaP-Agent0 powered by GPT-6 Astra Pro (60.0%). After a common calibration and hardware-adaptation procedure, the frozen system succeeds in all 30 physical trials, with ten trials on each of three tasks. Project Website: https://rpg-robot.github.io/

#robotics#manipulation#robot-learning
Read the original
See all →
Oct 31

Foundry Robotics raises $100 million Series A for autonomous assembly factories

Foundry Robotics announced an oversubscribed $100 million Series A funding round led by Marlinspike Partners. The round includes participation from early backers Khosla Ventures and Red Glass Ventures, alongside new investors SE Ventures, Wing VC, and Axiom Ventures.

Robotics & Automation News
Nov 2

Meta partners with Fourier on humanoid robot for household tasks

Meta and Chinese robotics startup Fourier jointly developed humanoid robot prototype Max, a 48-kilogram robot completed in three months. The companies are working on a successor model, targeting the household robotics market for housekeeping and daily tasks.

Korean outletAI Times
Nov 4
Paper

NEO-VLA, an encoder-free vision-language-action framework for robotic manipulation

The authors propose NEO-VLA, an encoder-free vision-language-action framework where images are tokenized and evolve jointly with language throughout the backbone. They show that this approach outperforms encoder-based counterparts across diverse manipulation tasks and distribution shifts.

Hugging FaceTrending 27
Oct 27
Paper

PDRVLA integrates physical dynamics into robot vision-language-action models

PDRVLA, a vision-language-action model with a lightweight physical dynamics module, generates robot actions from visual input and language instructions. It learns from public data augmented with human-to-robot synthesis and embodied chain-of-thought annotation, achieving 99.3% success on LIBERO and up to 93.6% on RoboTwin 2.0.

Hugging Face +1 moreTrending 25
Read the original