sayzardMin Choi (@minchoi)
휴머노이드 로봇이 불완전한 인간 동작 데이터를 바탕으로 스포츠 동작을 학습하기 시작했다는 소식입니다. 불완전한(노이즈 있는) 인간 모션 데이터를 이용해 실제 동작을 학습·적용하는 연구/응용으로, 로봇 제어와 모방학습 분야에서 의미 있는 진전으로 해석됩니다.
Ilir Aliu (@IlirAliu_)
완전 3D 프린트 가능한 저비용 오픈소스 로봇 팔이 공개되었습니다. 설계는 전체 3D 프린트 기반으로 7+1 DOF를 지원하며 물리적 AI 연구와 모방학습(imitation learning)에 적합합니다. 커스텀 하드웨어가 불필요해 프린트 후 서보 키트만 추가하면 되고, URDF·STEP·STL 파일이 제공되어 즉시 시뮬레이션·제작·연구에 활용할 수 있습니다.
https://x.com/IlirAliu_/status/2027306785757966368
#robotics #opensource #3dprinting #imitationlearning #hardware
Ilir Aliu (@IlirAliu_) on X
A highly affordable, fully 3D-printed robotic arm. [📍 open source] Features: > Fully 3D printed design > 7+1 DOF > or physical AI research & imitation learning Zero custom hardware needed, just print and add your servo kit! > URDF files > STEP & STL files >
Microsoft Research (@MSFTResearch)
Predictive Inverse Dynamics Models(PIDMs)이 모방학습(imitation learning)에서 표준 행동 복제(Behavior Cloning)보다 성능이 높은 이유를 분석한 연구 소개입니다. 다음 상태를 예측해 모호성을 줄이고 훨씬 적은 시연으로 학습할 수 있음을 보여줍니다. (원문 링크 제공)
Microsoft Research (@MSFTResearch) on X
This research looks at why Predictive Inverse Dynamics Models often outperform standard Behavior Cloning in imitation learning. By using simple predictions of what happens next, PIDMs reduce ambiguity and learn from far fewer demonstrations. Learn more: https://t.co/VTAklmQpcM
Rohan Paul (@rohanpaul_ai)
새 논문은 로봇이 시연을 단순히 모방하는 대신 도달 가능한 중간 목표(stepping stone goals)로 변환해 장기 과제를 학습할 수 있음을 보입니다. 제안된 방법 'Cago'(Capability Aware Goal Sampling)는 에이전트가 도달 가능한 서브골을 추적·선택하도록 하여 장기 과제 완료율을 높이고 기존 모방학습의 한계를 완화합니다.
Rohan Paul (@rohanpaul_ai) on X
This paper shows how robots can learn long tasks by turning demonstrations into reachable stepping stone goals instead of copying actions. Cago, short for Capability Aware Goal Sampling, helps agents finish long tasks by chasing reachable subgoals. Standard imitation learning
TeixiSo many gems in this interview, just little spoiler:
#DwarkeshPatel Next token prediction!
#RichardSutton That’s not a goal. It doesn’t change the world…
ps: My goal now is aging with such clarity thinking, and relaxed dialectical teaching!!!
#TheBitterLesson #RL #ML #LLMs #ImitationLearning #GoalDrivenExperience
Richard Sutton – Father of RL thinks LLMs are a dead end
Ars Technica NewsHow a big shift in training LLMs led to a capability explosion https://arstechni.ca/FerG #reinforcementlearning #imitationlearning #explainers #Features #AI
KINEWS24Nvdia and Stanford invent cross-episodic curriculum!
#Transformer #AI #DeepLearning #ReinforcementLearning #ImitationLearning #ContinuousControl #DiscreteControl #CurriculumLearning #FoundationModels #RoboMimic #efficiency #generalization
https://kinews24.de/nvdia-und-stanford-erfinden-cross-episodic-curriculum/
Nvdia und Stanford erfinden Cross-Episodic Curriculum - KiNews24.de
Nvdia und Stanford erfinden Cross-Episodic Curriculum: Forscher von Stanford, NVIDIA und der University of Texas stellen mit Cross-Episodic Curriculum (CEC) einen neuen KI-Algorithmus vor, der das Lernen und die Generalisierungsfähigkeit von Transformer-Agenten steigert.
JulianImitation learning is a powerful approach for training agents in sparse reward environments. By giving the agent a role model to learn from, we can help it to avoid making costly mistakes. But be careful not to be too perfect! The agent needs to learn to recover from its own mistakes. #ai #imitationlearning #reinforcementlearning #machinelearning
The GeekAnother step in AI development, thanks to Minecraft!
Stefan Ferreira 🏳️🌈OpenAI makes #imitationlearning breakthrough, creating the best #Minecraft bot yet by making it watch 70,000 hours of video of people playing the game.
#ai #technology