00:00
Dorsa Sadigh
577 posts
Dorsa Sadigh
@DorsaSadigh
CS Faculty @Stanford, @StanfordAILab, @StanfordHAI
Research scientist @GoogleDeepMind
PhD and BS from @Berkeley_EECS
- Congratulations to @ebiyik_ for successfully defending his dissertation, Learning Preferences for Interactive Autonomy! Erdem was my first student, and it has been wonderful working together!
- Excited to be awarded the AFOSR YIP this year :) Really looking forward to the project and better understanding conventions and trust in repeated human-robot interactions.Congratulations to this yearās Young Investigator Research Program (YIP) recipients! š @AFOSR will award grants to 36 scientists and engineers from 27 research institutions and businesses. @AFResearchLab #BasicResearch #Science #Engineering #AFOSRYIP afrl.af.mil/News/Article-Dā¦
- So it turns out fine-tuning miserably fails at times often due to the fact that the initial policy is in a different homotopy class than the final policy. We have a new fine-tuning algorithm Ease-In-Ease-Out that enables efficient transfer RL across homotopy classes.
GIF - Congratulations to Annie Xie, @loseydp, Ryan Tolsma, and @DorsaSadigh on the best paper award at @corl_conf! š It was a privilege to work with such an awesome group. Annieās talk is here: youtu.be/8bYv3BYmc8s?t=ā¦
- We solve RLHF by simply using supervised learning - w/o dealing with challenges of RL. Contrastive Preference Learning uses a regret-based preference model which not only accurately captures human preferences but also enables learning preferences in arbitrary MDPs. š§µšExcited to announced Contrastive Preference Learning (CPL), a simple RL-free method for RLHF that works with arbitrary MDPs and off-policy data. arXiv: arxiv.org/abs/2310.13639 With @rm_rafailov @harshit_sikchi @chelseabfinn @scottniekum W. Brad Knox @DorsaSadigh A threadš§µš
GIF - Beginning of many future exciting RoboNLP work. Learning adaptive language interfaces through interaction: arxiv.org/abs/2010.05190 w/ @siddkaramcheti and Percy Liang
- Bay Area Robotics Symposium (BARS) will be happening on Nov 20 this year: bars2020.github.io . We will have a series of faculty and student spotlight talks from Berkeley, UCSC, and Stanford. Also tune in for our keynote by @rodneyabrooks. (1/2)
- Imagine two robots helping you with a grocery bag. You can physically move the arms one at a time over a sequence of trajectory corrections. We'd like the robots to learn for once that the bag should be placed on the right while not stretching or squeezing the bag too much. (1/2)
GIF - So lucky to have such an awesome group of students. This was such a great surprise :) Thank you #ILIAD folks! @ebiyik94, @MinaeKwon, Mengxi Li, @zhangjie_cao, @siddkaramcheti, @andyshih_ , Suneel Belkhale, Megha Srivastava, Ian Huang, @woodywang153, @suvir_m, Vivek Myers,
- In large-scale pretraining for robotics -- where massive datasets are lacking -- data quality matters more than ever. We use mutual information estimators to identify high-quality data --optimizing for diverse states & easy-to-fit actions to improve learning policies.Behavior cloning... clones behaviors, so naturally data quality directly affects performance. However, there aren't great ways of measuring how "good" or "bad" different demonstrations are. Our recent work seeks to address this problem using estimators of mutual information... š§µ
GIF - Replying to @DorsaSadighFun fact that the robot and fox are completely OOD and are random toys that I grabbed from Niv's toy box!
- Congratulations to Hong Jun Jeon and @loseydp for being nominated for best student paper award at #RSS2020 for their work on Shared Autonomy with Learned Latent Actions.
- Access to internet-scale data is a luxury that robot learning lacks. This raises a need for developing data quality metrics that enable both curating offline datasets and guiding online data collection. (1/8)
