Giving history to our robot policies is crucial to solve a variety of daily tasks. However, diffusion policies get worse when adding history. 🤖. In our recent work we learn how adding an auxiliary loss that we name Past-Token Prediction (PTP) together with cached embeddings

Very happy to share that our work on learning long-history policies received the Best Paper Award from the Workshop on Learned Robot Representations @RoboticsSciSys ! 🤖🥳. Check out our paper if you haven't already! Thank you to all the organizers and

RT @KilianSFcat: I farem estada a Standford Califòrnia amb el @marceltornev

RT @liu_yuejiang: 🧠Memory is crucial for robots — to handle occlusions, track progress, stay coherent, etc. Yet, most VLA truncate context.….

Check out our paper and website for more details! And to see the set of real world tasks that we solve!. Really happy to be releasing my first paper of the PhD! Huge thanks to my co-leads @tangerinecoder and @liu_yuejiang and my advisor @chelseabfinn . Paper:

Finally, PTP enables a test-time scaling technique based on comparing your past rolled-out actions with your current predictions as a way of verifying that the policy is paying attention to the past. Check the paper for some promising preliminary results!

With the previous finding, to make the training more efficient we propose caching embeddings from the short-context history. We observe that we can train our policies much faster while keeping high performance.

2⃣ Efficiency: In our analysis we learn that we can reuse the encoder from a short-context policy in our long-history policy!

To bring back the correlation to long-history policies we add the PTP (Past-Token Prediction) auxiliary loss that consists of predicting the actions for each history timestep passed into the policy.

1⃣Efficacy: We ran an analysis on the effect of adding history in diffusion-based policies and observe a major difference compared to regression-based policies. For regression-based policies → history leads to high action correlation (copy-cat). For diffusion-based policies →

We identified two main challenges to give history to robot policies:. 1⃣ Efficacy: in the vanilla implementation when adding history the success rate of the policies drop significantly. 2⃣ Efficiency: when increasing context GPU memory consumption increases making the training.

This is very very cool!! You should try it out! Congrats to @sequoia on making the great decision to back @itsalfredw and @florian_jue !.

RT @itsalfredw: AI writes your code. Now it talks to your users. We raised $27M from @Sequoia to build @ListenLabs. Listen runs thousands….

RT @_anniechen_: Not all human-collected demos are created equal:.✔️ All are successful.❌ But some strategies are unreliable or brittle. Th….

RT @tangiblerobots: Robots can do flips and play chess, but they still can’t grab a snack or clean your table. Teleoperation is the key to….

RT @pulkitology: Overcoming the lack of reliability of Behavior cloning (BC) with reactive reinforcement learning. Action-chunking is a tw….

RT @abhishekunique7: So I heard we need more data for robot learning :) Purely real world teleop is expensive and slow, making large scale….

Check out the paper! And the USDZ assets that we open sourced!. Paper: Website: USD assets: Thank you to the amazing team who made this possible @prodarhan, @carrieyuanjiayi, Macha, @larsankile,.

CASHER gives us a glimpse on how to solve the data scarcity problem in robotics. We see promise on how to scale up the amount of data obtained while reducing the human effort through time and making the robotics data flywheel a reality. We solve the environment design challenge

The same system presented in CASHER has the capability of fine-tuning generalist policies in simulation without needing to collect any additional human data! We rollout the policy in simulation to collect a few successful trajectories and we finetune it with RL from the