RT @Haoyu_Xiong_: Your bimanual manipulators might need a Robot Neck 🤖🦒. Introducing Vision in Action: Learning Active Perception from Huma….

So excited for @siddk to begin this new chapter! Working with Sidd has been a blast -- one of the most brilliant and kindest people I have had the privilege to learn from. Future PhD students, apply to work with Sidd!!.

RT @jenngrannen: Meet ProVox: a proactive robot teammate that gets you 🤖❤️‍🔥. ProVox models your goals and expectations before a task start….

RT @oshaikh13: What if LLMs could learn your habits and preferences well enough (across any context!) to anticipate your needs?. In a new p….

RT @verityw_: Embodied chain-of-thought reasoning (ECoT) is a powerful way to improve robot generalization & performance. But why is this t….

RT @perryadong: Robotic models are advancing rapidly—but how do we scale their improvement? 🤖. We propose a recipe for batch online RL (tra….

Data quality can have a big impact on the performance of behavior cloning methods. But how can we measure the quality of demonstrations? One way is to score demonstrations via mutual information estimators:. Check out @JoeyHejna's thread 👇.

Human video data can be easier to collect than robot demonstrations—but extracting actions for training robot policies is challenging. RAD uses language reasoning extracted from human videos to boost generalization in reasoning-based policies. Check out @jadenvclark's thread 👇.

RT @siddkaramcheti: Really grateful to @StanfordHAI for covering our work on Vocal Sandbox - a framework for building robots that can seam….

This was a fun project to be a part of! Led by @jenngrannen and @siddkaramcheti (on the faculty job market!); w/ @percyliang and @DorsaSadigh. Paper: Website:

Introducing Vocal Sandbox 💬🤖: a framework for enabling human-robot collaboration by building up & adapting robot behaviors at different levels of abstraction via multiple teaching modalities. Check out @jenngrannen's thread below 👇.

This was a fun project with David Yuan (undergrad!), @kaylburns , Sazzad Islam (undergrad!), @tonyzzhao, @chelseabfinn, and @DorsaSadigh. ArXiv: Website: (5/5).

But learning from mixed-quality crowdsourced data also presents challenges and opportunities for future work! We’ll be presenting more about RoboCrowd at the Workshop on Mastering Robot Manipulation in a World of Abundant Data @corl_conf on Saturday 11/9. (4/5)

We validate various incentive mechanisms for shaping user interactions into useful data—and further, see evidence that the crowdsourced data can be useful for pre-training policies that are then fine-tuned on expert data. (3/5)

We instantiate the framework in a system based on ALOHA (@tonyzzhao et al. 2023). We deployed the system for two weeks near a cafe on Stanford campus. Over 200 users voluntarily came up to the robot and interacted with it, giving us over 800 episodes (2/5)

We propose RoboCrowd, a framework for collecting robot demonstrations via crowdsourcing, gamification, and incentive design. Our key idea is to align user incentives like material rewards 🍬, intrinsic interest 🤔, and social comparison 🎮 with getting useful robot data 🧵(1/5)

Thanks to great collaborators: @suneel_belkhale , @JoeyHejna , Evelyn Choi (undergrad!), Sazzad Islam (undergrad!), and @DorsaSadigh. ArXiv: Website: (7/7).

We’re presenting our “negative results” at @corl_conf -- stop by our poster in Session 4 (4:00-5:30pm) this Friday 11/8. (6/7)

And even on simpler tasks in real/simulation with minimal environment challenges, we found across a variety of methods that autonomous IL could be mildly helpful to performance, but reallocating our effort to collecting more data for IL was often simpler and more efficient. (5/7).

In our experience: not really. We document a variety of challenges in setting up *truly autonomous* data collection. Addressing these challenges required effort on either instrumenting environments or shaping tasks to become less useful/realistic. (4/7)