Chelsea Finn

@chelseabfinn105k followers

Asst Prof of CS & EE @Stanford Co-founder of Physical Intelligence @physical_int PhD from @Berkeley_EECS, EECS BS from @MIT

Open on X

Top outlier

41.7×

Best organic post

Peak views

615k

Single post

Bookmarks

13k

Across 98 organic posts

Last scan

Oct 17

Community cache

Posts 98 posts

#	Date	Post	↓ Outlier	Views	Likes	Reposts	Bookmarks
1	Apr 1, '23	In light of tremendous AI advances & recent calls for pauses on AGI research, I've decided to pivot our lab's direction. Going forward, we’ll focus entirely on certain real-wor	41.7×	615k	994	68	227
2	Oct 31, '24	Very happy to start sharing our work at Pi 🤖❤️ - a 3B pre-trained generalist model trained on 8+ robot platforms - a post-training recipe that allows robots to do dexterous, long-	24.4×	360k	1.9k	251	772
3	Mar 27, '23	We introduce a system for fine-grained robotic manipulation! 🤖 What’s new? * We can control cheap robots to do surprisingly dexterous tasks * New technique that allows robots to l	17.9×	265k	947	189	245
4	Jan 27, '23	LLMs like ChatGPT are becoming more fluent – how can we detect if something was written by a language model or a human? We developed DetectGPT: a method for detecting if a passage	16.8×	248k	1k	204	309
5	Jul 27, '23	Our robot can now make you coffee 🤖☕ A short 🧵 on how it works ⬇️	14.2×	210k	855	132	144
6	Feb 25, '20	Want to learn about meta-learning? Lecture videos for CS330 are now online! Topics incl. MTL, few-shot learning, Bayesian meta-learning, lifelong learning, meta-RL & more: + 3	13.9×	0	2.5k	655	599
7	Mar 12, '24	I’m really excited to be starting a new adventure with multiple amazing friends & colleagues. Our company is called Physical Intelligence (Pi or π, like the policy). A short th	12.3×	182k	1.6k	103	320
8	Jan 23, '25	Disappointed with your ICLR paper being rejected? Ten years ago today, Sergey and I finished training some of the first end-to-end neutral nets for robot control 🤖 We submitted th	11.5×	170k	1.6k	154	302
9	Feb 1, '23	Want to try out DetectGPT? We just released a demo — try it out here: See if you can fool it and let us know your feedback.	11.4×	168k	618	120	255
10	Apr 17, '23	Want to learn about meta-learning & few-shot learning? All of the latest lecture videos for Stanford CS330 are now online! New topics in Fall '22 include: - self-supervised pre	10.2×	151k	910	184	533
11	Dec 12, '23	DPO is a runner up for NeurIPS outstanding paper. 🙌 Big congrats especially to the students @rm_rafailov @archit_sharma97 @ericmitchellai & the other awardees. If you haven't	10×	148k	839	84	257
12	Jan 3, '24	Meet Mobile ALOHA! 🤖 What’s new: - low-cost but widely-capable teleop platform 🦾🛞 - imitation learning just works Paper/Code/Videos: Led by @zipengfu and @tonyzzhao	9.3×	138k	731	130	175
13	Mar 5, '23	Can you peel an egg without tactile feedback? 🥚 Turns out yes! We had some fun pushing the limits of our robot hardware last night. 😄🤖	8.2×	120k	627	69	46
14	Apr 1, '23	The lab's central goal will be to improve eye care for amphibians with robots. See ⬇️ for our initial work: fitting contact lenses for toads. We hope their improved vision will hel	8.0×	119k	370	25	12
15	Dec 6, '23	LLMs fine-tuned with RLHF are known to be poorly calibrated. We found that they can actually be quite good at verbalizing their confidence. Led by @kattian_ and @ericmitchellai,	7.5×	110k	349	47	189
16	Mar 20, '24	Introducing a new, fully open robotics dataset! - 76k episodes - 564 unique scenes - 100 contributors - 13 labs/institutions - 3 continents A short 🧵 on the backstory	6.7×	99k	873	142	421
17	Feb 4, '25	Our first open-source release at Pi 🤖 - π₀ and π₀-FAST model weights - code for model, on-robot inference, & fine-tuning - fine-tuned checkpoints that can be used out-of-the-b	6.1×	90k	778	105	210
18	Oct 3, '23	What happens when 100+ robot learning researchers share data? ⛁ 1M robot demos open-sourced in a common format 🤖 policies trained on this data can control multiple robot embodim	6.1×	89k	481	98	111
19	Sep 26, '19	Excited to share that I’m teaching a new course on multi-task & meta-learning! Topics incl. optimization-based meta-learning, lifelong learning, meta-RL , etc Slides & as	5.7×	0	1.3k	233	158
20	Jul 24, '23	Language models are trained to directly predict text. Can AI agents learn to read text without explicit language supervision? 🧵 on our #ICML2023 paper led by Evan Liu. ↓	5.2×	77k	347	70	170
21	Nov 18, '25	For robots to be actually useful, they need to be reliable. We’re sharing an RL recipe for VLA models that takes a step in this direction, allowing robots to operate autonomously f	5.1×	75k	827	104	314
22	Oct 4, '25	RL fine-tuning often prematurely collapses policy entropy. We consider a general framework, called set RL, i.e. RL over a set of trajectories from a policy. We use it to incentiviz	5.0×	74k	741	63	550
23	Nov 8, '24	At @corl_conf, I’m giving a talk on π₀ post-training, i.e. how we got the robot to fold laundry. Tomorrow/Saturday at 1:45 pm at the WCBM workshop:	4.6×	68k	273	32	88
24	Jun 29, '18	I’m excited to share that I will be joining the faculty of @Stanford in Computer Science, starting in 2019!	4.4×	0	1.4k	44	5
25	Oct 4, '25	Hierarchical RL for LLM reasoning. Paper:	4.3×	63k	592	64	412
26	Jul 8, '20	Convolution is an example of structure we build into neural nets. Can we _discover_ convolutions & other symmetries from data? Excited to introduce: Meta-Learning Symmetries by	4.2×	0	824	168	185
27	Oct 24, '25	Ctrl-World is a controllable world model that generalizes zero-shot to new environments, cameras, and objects. Paper: Model & code: The results are exciting — a short thread on	4.2×	62k	359	37	214
28	Oct 27, '22	Common fine-tuning wisdom is to adapt the last layer or the entire neural net. We find that, sometimes, fine-tuning only the first layers or middle layers works best. Paper: A sh	4.1×	0	784	124	230
29	Jun 14, '24	Really excited to share OpenVLA! - state-of-the-art robotic foundation model - outperforms RT-2-X in our evals, despite being nearly 10x smaller - code + data + weights open-source	4.1×	61k	282	55	108
30	Aug 31, '23	In robotics, a key bottleneck is data. We don’t have the ImageNet or Wikipedia for motor control. 🤖 If we can use data from multiple robot platforms, we can overcome this bottle	4.1×	61k	279	45	63
31	Apr 17, '24	Excited to share one of the last projects I worked on at Google -- We picked three of the most dextrous tasks we could think of: - tying shoelaces 👟 - replacing a robot finger 🤖	3.8×	56k	407	42	73
32	Nov 3, '23	The Bay Area Robotics Symposium (BARS) last week was a success! All of the talks, including keynotes from @SebastianThrun and @jon_barron, are now available online:	3.8×	56k	263	32	64
33	Oct 23, '25	Flow matching for value functions. Works well for both offline and offline-to-online RL. Paper:	3.7×	55k	532	56	304
34	Sep 8, '25	π0.5 weights + fine-tuning code are now open-source!	3.6×	53k	405	35	124
35	Jan 4, '24	More examples of everything that Mobile ALOHA can do when human-teleoperated. ⬇️ Fortunately, Zipeng took the robot home on laundry day.	3.5×	52k	301	41	39
36	Feb 27, '23	Turns out NeRFs are super useful for robot grasping! We use NeRFs for data augmentation, for imitation learning with wrist cameras. -> Leads to >20% higher success rates. Wit	3.5×	52k	231	41	31
37	Oct 27, '25	Most VLAs have no memory. MemER is a scalable method for giving VLAs long-term visual memory. Key ideas: - relevant info is in a small # of frames: store just a few key frames - ma	3.4×	50k	454	45	271
38	Jul 28, '23	Vision-language ➡️ vision-language-action model By using a pre-trained VLM (e.g. PaLI-X), RT-2 enables robots to generalize to new objects & instructions RT-2 also shows basic	3.4×	50k	245	55	52
39	Sep 12, '23	Meet our new quadruped! 🤖🐕 1. Train behaviors in simulation with a curriculum. 2. Distill into single depth-based policy. 3. Deploy on the real robot. Paper & Videos: Code:	3.3×	49k	246	42	39
40	Feb 17, '24	Want to optimize/edit/transform neural net weights using another neural net? Universal neural functionals can be applied to any architecture, while appropriately handling its permu	3.2×	47k	250	29	124
41	Apr 22, '25	Introducing π-0.5! The model works out of the box in completely new environments. Here the robot cleans new kitchens & bedrooms. 🤖 Detailed paper + videos in more than 10 unse	3.2×	47k	700	96	187
42	Oct 22, '21	Large language models (LLMs) often make mistakes that are difficult to correct. We study the problem of quickly editing these models: Paper: Code: w/ @_eric_mitchell_, C. Lin, @ABo	3.0×	0	534	116	147
43	Jul 7, '20	Supervised ML methods (i.e. ERM) assume that train & test data are from the same distribution, & deteriorate when this assumption is broken. To help, we introduce adaptive	2.9×	0	571	106	139
44	Jun 27, '18	I’m thrilled and honored to have received the MIT TR 35 innovator award this year. Grateful to have been able to work with some amazing faculty, students, and researchers thus far	2.9×	0	830	64	10
45	Feb 9, '22	What should ML models do when there's a perfect correlation between spurious features and labels? This is hard b/c the problem is fundamentally _underdefined_ DivDis can solve th	2.9×	0	539	87	167
46	Feb 29, '24	The standard way to supervise ML models it to give many many examples. We can provide supervision far more efficiently: 1. verbally describe how the model fails 2. translate feed	2.7×	40k	214	42	114
47	Jan 16, '25	New way to predict robot actions! 🤖 Frequency-space action tokenizer (FAST) enables LLM-like token prediction for dexterous robot actions When we train on the DROID dataset, the r	2.6×	39k	298	45	110
48	Feb 28, '25	We can now fine-tune the 7B OpenVLA model for 50 Hz control! 🤖 Paper: Code:	2.6×	39k	402	55	165
49	Jul 28, '25	EXPO is a new method for RL post-training of diffusion policies. With good pre-training, it’s both really stable and ~2x more efficient than the most efficient RL fine-tuning metho	2.6×	39k	367	55	216
50	May 15, '19	I am honored to have received the ACM Doctoral Dissertation Award this year. Thank you to my amazing advisers, mentors, and collaborators at @UCBerkeley & @GoogleAI, especially	2.5×	0	803	36	3
51	Jul 20, '21	Does your neural network struggle with spurious correlations? Check out Evan’s long talk at #ICML2021 on why they should just train twice (JTT). Paper: Talk: Code:	2.5×	0	411	82	158
52	Jan 22, '25	Why is action chunking crucial for robot dexterity? 🤖 - We identify a natural tradeoff between temporal consistency and reactivity - New policy decoding technique that is both t	2.4×	36k	335	60	179
53	Sep 11, '19	It's hard to scale meta-learning to long inner optimizations. We introduce iMAML, which meta-learns without differentiating through the inner optimization path using implicit dif	2.4×	0	513	116	61
54	Dec 1, '18	Excited to share our most recent work! We train a single world model from unsupervised interaction and use that model to perform many different tasks in the real world, from raw pi	2.3×	0	522	127	27
55	Feb 8, '24	ALOHA 2 is here! 🤖 The upgraded hardware has: * More durability * Greater fine-grained control * Lower stress on operator All off-the-shelf + open-source.	2.3×	34k	254	43	62
56	Feb 26, '25	Can we prompt robots, just like we prompt language models? With hierarchy of VLA models + LLM-generated data, robots can: - reason through long-horizon tasks - respond to variety o	2.2×	33k	454	67	167
57	May 28, '20	Offline RL may make it possible to learn behavior from large, diverse datasets (like the rest of ML). We introduce: MOPO: Model-based Offline Policy Optimization w/ Tianhe Yu, Garr	2.2×	0	435	93	88
58	Jul 1, '25	We still lack a scalable recipe for RL post-training seeded with demonstration data. Many methods add an imitation loss, but this constrains learning too much. We propose to use th	2.2×	32k	313	44	212
59	Sep 26, '25	I’m giving two talks about new research at @corl_conf tomorrow! 🤖 - giving VLAs long-term memory (12 pm, RemembeRL workshop) - the bottlenecks for developing better robot foundati	2.1×	31k	338	17	87
60	Jan 22, '20	Excited to share PCGrad, a super simple & effective method for multi-task learning & multi-task RL: project conflicting gradients On Meta-World MT50, PCGrad can solve 2x	2.1×	0	405	97	75
61	Oct 5, '18	CACTUs: an unsupervised learning algorithm that learns to learn tasks constructed from unlabeled data. Leads to significantly more effective downstream learning & enables few-s	2.1×	0	439	113	45
62	Jun 27, '20	How can robots learn in changing, open-world environments? We study: Deep Reinforcement Learning amidst Lifelong Non-Stationarity with Annie Xie, @jmes_harrison @StanfordAILab (1/5	1.9×	0	369	86	79
63	Oct 19, '22	Can robots adapt on the fly when deployed? Our paper studies single life RL, where an agent must adapt a solve a new scenario in just one episode. #NeurIPS2022 paper, led by @_an	1.8×	0	410	50	80
64	Jul 15, '22	Neural nets are brittle under domain shift & subpop shift. We introduce a simple mixup-based method that selectively interpolates datapts to encourage domain-invariance ICML 22	1.8×	0	334	57	108
65	May 4, '20	Working on multi-task learning? Trying out PCGrad now requires only 1 more line of code. PyTorch version coming soon.	1.7×	0	356	66	63
66	Jul 20, '21	Thrilled to share new work on AI for education: can we give detailed, high-quality feedback to students? Post: NYT Coverage: A collab w. the amazing @mike_h_wu @chrispiech & co	1.7×	0	305	74	70
67	Sep 20, '22	Why are the motor skills of a toddler so hard for robots to develop? 🤖 @WIRED challenged me to explain Moravec’s Paradox at five levels of difficulty. A fun and accessible video,	1.7×	0	361	48	68
68	Feb 3, '22	Can robots 🤖 do tasks they weren't trained to do? BC-Z trains a robot to do 100 tasks prompted with language instructions. The robot generalizes to 20 new tasks. Our @GoogleAI B	1.6×	0	333	63	51
69	Feb 6, '18	Enabling a robot learn from a video of a human was a goal of mine 3.5 yrs ago. Now, we have! (in a limited sense) One-shot Imitation from Humans via Domain-Adaptive Meta-Learning K	1.6×	0	366	101	0
70	Jul 3, '19	How should robots learn from humans? While most use rewards or demos, we study how we might have agents learn tasks via interaction with humans. Try for yourself here! w/ Mark Wood	1.6×	0	313	91	39
71	Jul 24, '18	Students at MILA released a clean pytorch implementation of MAML for RL:	1.6×	0	336	85	36
72	Oct 25, '19	Tired of your robot learning from scratch? We introduce RoboNet: a dataset that enables fine-tuning to new views, new envs, & entirely new robot platforms. w/ Dasari @febert888	1.6×	0	331	85	35
73	Mar 21, '19	We introduce PEARL, a new meta-reinforcement learning method that is 20-100x faster and has better end performance than the best prior method. I'm excited to see what this method e	1.6×	0	337	82	35
74	Apr 16, '20	Honored to be selected for the MSR Faculty Fellowship award, which reflects not only on my work but also my students, collaborators, and advisers. Thank you Microsoft, and congratu	1.5×	0	493	12	7
75	Mar 6, '20	Imitation learning needs human demos to improve. In MILI, robots improve themselves w/ multi-task self-imitation Scalable Multi-Task Imitation Learning with Autonomous Improvement	1.5×	0	286	94	37
76	Sep 11, '19	Evaluating multi-task & meta RL at scale is hard b/c it needs broad + structured set of tasks. Introducing Meta-World Benchmark including 50 simulated manipulation tasks Pape	1.5×	0	316	77	32
77	Mar 15, '21	Predicting video is a powerful self-supervised approach for robots to learn about the world Greedy hierarchical VAEs can train larger models with less memory surprisingly outperf	1.5×	0	285	60	64
78	Mar 24, '22	Excited to share new work on pre-training a single reusable representation for many robot manipulation domains & tasks R3M pre-trains on human videos & then learns w/ only	1.4×	0	302	52	50
79	Dec 10, '19	Meta-learning has a peculiar, widespread problem that leads to terrible performance when faced with seemingly benign changes to the training set-up. We analyze this problem & p	1.4×	0	281	69	42
80	Dec 7, '21	Can NNs learn and enforce symmetries on their predictions? Inspired by Noether’s Thm, we do so by learning quantities that should be conserved. This method can recover known conser	1.4×	0	256	49	74
81	Sep 13, '21	If you have a lot of ML tasks & are not sure which are complementary: We introduce an efficient and effective approach for determining task groupings. With Chris Fifty, @esiami	1.3×	0	261	50	64
82	Jun 19, '19	Can the compositionality of language help RL agents solve long-horizon tasks? We develop an open-source CLEVR-like RL env, evaluate long-horizon reasoning + systematic generaliza	1.3×	0	293	58	28
83	Apr 7, '20	Supervising RL is hard, especially if you want to learn many tasks. To address this, we present: Weakly-Supervised Reinforcement Learning for Controllable Behavior with @rl_agent,	1.3×	0	265	58	45
84	Mar 5, '20	Robots can use meta-learning to quickly adapt from simulation to various real world conditions. Rapidly Adaptable Legged Robots via Evolutionary Meta-Learning with Xingyou Song, Yu	1.3×	0	284	61	33
85	Oct 31, '20	Reinforcement learning can lead to behaviors that don’t generalize. In our #NeurIPS2020 paper, we introduce a simple idea to allow extrapolation to new envs w/ a few trials: One So	1.3×	0	296	38	42
86	Jan 17, '18	The video from our ICML tutorial (w/ Sergey Levine) on Deep Reinforcement Learning, Decision Making, and Control is now online!	1.3×	0	287	91	2
87	Mar 2, '21	Last week, I gave a talk ‘at’ Toronto discussing 3 principles for tackling distribution shift: * pessimism * adaptation * anticipation and how they can improve robustness to spurio	1.3×	0	250	40	70
88	Dec 2, '18	@ankesh_anand @egrefen @ylecun I posted the slides for my model-based RL tutorial today here: Content is similar to the DRL bootcamp video 1 yr ago, but updated.	1.2×	0	242	66	33
89	Apr 27, '25	I'm giving two talks today/Sunday at #ICLR2025 ! - Post-Training Robot Foundation Models (Robot Learning Workshop @ 12:50 pm) - Robot Foundation Models with Open-Ended Generalizati	1.2×	18k	239	11	45
90	Apr 17, '19	End-to-End Robotic Reinforcement Learning without Reward Engineering: Our latest work on enabling robots to learn skills end-to-end from pixels in 1-4 hrs. Paper: Code: w/ @avising	1.2×	0	263	64	14
91	Dec 22, '20	Interested in offline RL? Handling image observations & continuous actions is important for offline RL in the real world. We introduce LOMPO to tackle this setting. with Rafael	1.2×	0	252	44	41
92	May 8, '19	If you weren't able to make it to ICLR, my talks (& others!) are: What can we learn from unlabeled interaction? slides: video: Meta-Learning: Challenges & Frontiers slides:	1.2×	0	259	51	29
93	Sep 15, '19	New paper: Hierarchical visual foresight learns to generate visual subgoals to break down a goal into smaller pieces. Accomplishes long-horizon vision-based tasks, without any su	1.2×	0	260	45	35
94	Dec 15, '20	Interested in distribution shift & ML in the real world? Excited to share WILDS: We talked to domain experts & put together a benchmark of real-world ML problems that all e	1.1×	0	251	40	39
95	Oct 23, '20	Can robots learn to autonomously explore their environment? We introduce Batch Exploration with Examples (BEE) led by @anniee268, Alex Nam, & @SurajNair_1 Thread👇 (1/8)	1.1×	0	248	40	40
96	May 2, '20	Interested in unsupervised meta-RL, where agents propose their own tasks and learn how to learn them? Check out a new blog post by @abhishekunique7 and Ben Eysenbach:	1.1×	0	263	34	31
97	Jun 9, '19	At @icmlconf and interested in how we can enable more flexible neural nets that can build upon previous data? I'm giving a tutorial on meta-learning for fast adaptation with @svlev	1.0×	0	255	33	18
98	Jun 17, '19	Great to see lots of interest in meta-learning at #CVPR2019 ! Had trouble getting in the room to give my talk. My talk was based on our @icmlconf tutorial with @svlevine Slides, vi	1.0×	0	238	42	13
99	Nov 18, '20	Congratulations to Annie Xie, @loseydp, Ryan Tolsma, and @DorsaSadigh on the best paper award at @corl_conf! 🎉 It was a privilege to work with such an awesome group. Annie’s talk	1.0×	0	258	16	20
100	Aug 4, '22	Congratulations to @TianheYu who defended his PhD thesis this week!👏🎓 His work includes: - Meta-World - offline model-based RL methods like MOPO and COMBO - methods for using unl	0.9×	0	256	14	21

Scan another creator

Run this on any account.

XOutlier builds this table for every public profile you open on x.com, with live data and every post, not just the top 50.

Add to Chrome Download .zip