Idanwo ti o sopọ mọ itọsọna · Alabọde Ipele
Awọn awoṣe Iṣe-Ede-iriran fun Idanwo Robotics
Learn how vision-language-action models combine camera input and instructions with robot commands, and where their evidence comes from.
Ibeere 1 ti 8
Awọn oriṣi mẹta ti igbewọle/jade ti n ṣalaye awoṣe Iṣe-ede-Iran (VLA) kan?
Awotẹlẹ gbogbo ibeere8 ibeere
Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.
- Awọn oriṣi mẹta ti igbewọle/jade ti n ṣalaye awoṣe Iṣe-ede-Iran (VLA) kan?
- Awọn aworan kamẹra, itọnisọna ede, ati awọn iṣe robot
- Ohùn, ọrọ, ati awọn aworan
- Awọn akọọlẹ sensọ, GPS, ati ipo batiri
- Koodu, igbejade alakojo, ati awọn abajade idanwo
- Bawo ni Google DeepMind's RT-2 ṣe aṣoju awọn iṣe robot ki ẹrọ iyipada le ṣe agbejade wọn?
- Bi lemọlemọfún lilefoofo-ojuami fekito nikan
- Gẹgẹbi awọn ami iṣe ti a sọ di mimọ, ti o jọra si awọn ami ọrọ
- Bi ROS ifiranṣẹ XML
- Gẹgẹbi awọn iye ẹsan fun ẹkọ imuduro
- Kini 'gbigbe pajawiri' tumọ si ni ọrọ ti awọn VLA?
- Robot naa gbona pupọ o si tiipa
- Awọn roboti meji ṣe paarọ sọfitiwia wọn
- Imọ iwọn wẹẹbu lati ikẹkọ iṣaaju ṣe iranlọwọ fun robot mu awọn nkan aramada ati awọn ilana
- A ṣe iṣiro awọn iṣe laisi eyikeyi nẹtiwọọki nkankikan
- Kini anfani bọtini ti pi-0's itankale / ori iṣe ibamu-sisan ni akawe si awọn ami iṣe adaṣe ọtọtọ kan?
- O yọ awọn nilo fun eyikeyi awọn kamẹra
- O ṣiṣẹ nikan ni kikopa
- O ṣe imukuro itọnisọna ede
- O ṣe agbejade didan, awọn ṣoki iṣe igbohunsafẹfẹ-giga fun iṣakoso dexterous diẹ sii
- Kini idi ti Šii X-Embodiment datasetset ṣe pataki fun awọn VLA?
- O ṣe adagun data robot kọja ọpọlọpọ awọn oriṣi roboti lati jẹ ki gbogboogbo-iṣaro-agbelebu ṣiṣẹ
- O jẹ ala-ilẹ nikan fun ifori aworan
- O tọju data robot kan ṣoṣo
- O ti wa ni a ede-nikan ọrọ koposi
- Which data sources did DeepMind combine when co-fine-tuning RT-2?
- Only a fixed list of hand-written joint commands
- Robot demonstration data and web data with a pretrained vision-language model
- Only audio recordings without robot observations
- Only random action tokens with no examples
- What kinds of commands can an RT-2-style action representation encode?
- Only natural-language explanations with no robot controls
- A fixed video caption for each camera frame
- Discretized robot motion and gripper commands represented as tokens
- A robot’s hardware serial number
- How broad was the RT-2 evaluation of emergent capabilities reported by Google DeepMind?
- Exactly 12 trials
- One trial for every robot type worldwide
- No physical robot trials
- More than 6,000 robotic trials