Envisioned by many, brain-robot interface (BRI) stands out as a thrilling but challenging research topic.
It is an exciting time for BRI research. Brain signal decoding and robot intelligence are improved a lot by modern machine learning algorithms.
NOIR is a general-purpose, intelligent BRI system that enables humans to command robots to perform 20 challenging everyday activities using their brain signals, such as cooking, cleaning, playing games with friends, and petting a (robot) dog.
With 10-minute calibration for each session, 3 human participants successfully accomplished 20 long-horizon tasks (4-15 subtasks). On average, each task requires 1.8 attempts to succeed with an average task completion time of 20 minutes.
NOIR uses non-invasive EEG devices to record brain activities. We decode human intention, including what object to interact with (via SSVEP), how to interact (motor imagery), and where to interact (motor imagery).
The effectiveness of NOIR is improved by few-shot robot learning algorithms that are based on foundation models. This allows NOIR to adapt to individual users, predict their intentions, and reduce human time and effort.
NOIR holds a significant potential to augment human capabilities and enable critical assistive technology for individuals who require everyday support. We hope NOIR paves the path for future BRI research!
Decoded human intention signals are communicated to our robots. These robots are equipped with 14 pre-defined parameterized primitive skills, such as Pick(object,x,y,z).
Do you want to learn to train and evaluate embodied AI solutions for 1000 household tasks in a realistic simulator? Join our BEHAVIOR Tutorial at
#ECCV2022
: Benchmarking Embodied AI Solutions in Natural Tasks!
Time: Monday, Oct 24th 14:00 local time (4:00 Pacific Time)
How do humans transfer their knowledge and skills to artificial decision-making agents?
What kind of knowledge and skills should humans provide and in what format?
@RuohanZhang76
, a postdoc at
@StanfordSVL
and
@StanfordAILab
, provides a summary:
👇
We invite you to submit papers (up to 9 pages for long papers and up to 5 pages for short papers, excluding references and appendix) in the NeurIPS 2022 format. All submissions will be managed through OpenReview submission website. See Call For Papers 4/N
How to harness foundation models for *generalization in the wild* in robot manipulation?
Introducing VoxPoser: use LLM+VLM to label affordances and constraints directly in 3D perceptual space for zero-shot robot manipulation in the real world!
🌐
🧵👇
In honor of our upcoming
@NeurIPSConf
workshop on "All Things Attention", and the fact that the deadline for you to submit your work has been extended to **Oct 3**, I present a thread on attention and decision making in AI!
The Submission Deadline has been extended to Oct 3, 2022 (11:59PM AoE)
@NeurIPSConf
Consider submitting your work to our workshop
@attentioneurips
See details here:
Robot visual navigation in unseen homes is hard: end-to-end RL works well in sim but gets only 23% real-world success.
Today, in the first real-world empirical study of visual navigation, we show Modular Learning achieves 90% success in unseen homes!
1/N
How can we effectively predict the dynamics of multi-agent systems?
💥 Identify the relationships. 💥
We are excited to share IMMA at
#Neurips2022
, a SOTA forward prediction model that infers agent relationships -- simply by observing their behavior.
1/
@chenwang_j
it’s a pleasure to work with you and this team. The key insight is that for robot learning from humans, data for training high-level planer and low-level visuomotor skills can be different. PLAY data is a good candidate for learning to plan.
How to teach robots to perform long-horizon tasks efficiently and robustly🦾?
Introducing MimicPlay - an imitation learning algorithm that uses "cheap human play data". Our approach unlocks both real-time planning through raw perception and strong robustness to disturbances!🧵👇
Our language decoding paper (
@AmandaLeBel3
@shaileeejain
@alex_ander
) is out! We found that it is possible to use functional MRI scans to predict the words that a user was hearing or imagining when the scans were collected