Olivia Lee

AI @Figure_robot, building Helix | Prev. Robot Learning @StanfordAILab @StanfordIPRL

[Olivia Lee @olivia_y_lee]

We have the robots, the data, and the compute. It’s time to scale up.

Enjoy 4 hours of Helix 2.5 working in unseen homes with unseen objects.

[Brett Adcock @adcock_brett]

Here's 4 hours of our robot doing zero-shot work in 30 rental homes

[Olivia Lee @olivia_y_lee]

@leto__jean Thanks Jeannette! I would love to catch up soon

[Olivia Lee @olivia_y_lee]

We have the robots, the data, and the compute. It’s time to scale up.

Enjoy 4 hours of Helix 2.5 working in unseen homes with unseen objects.

[Brett Adcock @adcock_brett]

Here's 4 hours of our robot doing zero-shot work in 30 rental homes

[Olivia Lee @olivia_y_lee]

I am so thrilled to finally share this with the world!

Helix 2.5 is just the beginning of scaling robot learning to more environments and tasks.

Read more on our blog: https://t.co/yfeBIlw6qZ (4/4)

[Olivia Lee @olivia_y_lee]

Presenting Helix 2.5, Figure’s most advanced neural network.

Helix 2.5 performs three whole-body long-horizon tasks across 30 unseen homes.

To our knowledge, this is the first instance of zero-shot whole-body generalization at this scope on a humanoid.

(1/4)

[Figure @Figure_robot]

Today we’re releasing Helix 2.5

We rented 30 homes in the Bay Area. The robots arrived with no additional training and started doing useful work

[Olivia Lee @olivia_y_lee]

We achieve this by pretraining on Index, Figure’s global-scale dataset of human behavior.

Our key finding: Index pretraining accounts for most of Helix 2.5's zero-shot capability.

Whole-body intelligence is learned from human experience and transferred to new scenarios. (3/4)

[Brett Adcock @adcock_brett]

We just wrapped what began as an 8-hour challenge - and it ran for 200 hours without a failure

Shoutout to the team for the hardcore engineering behind F.03 and the robust Helix models powering it

[The Humanoid Hub @TheHumanoidHub]

Figure just set a new standard in unedited humanoid demos.

They completed 24 hours of autonomous work on a live stream, where three Figure 03 robots took turns processing delivery packages.

At a rate of ~21 packages/minute, that is less than 3 seconds per package.

[Figure @Figure_robot]

Today we're showing Helix 02 that can tidy a living room fully autonomously

Figure is designed so when you leave the house, your home resets exactly how you like it

[Olivia Lee @olivia_y_lee]

@tylerlum23 @kushalk_ So cool! Congrats on the work @tylerlum23 @kushalk_

[The Humanoid Hub @TheHumanoidHub]

Figure 03 demonstrating autonomous coordinated bimanual dexterity with the help of tactile sensing and palm cameras:

- Unscrewing a bottle cap
- Picking a pill from a medicine box
- Dispensing exactly 5 ml from a syringe
- Sorting metal pieces

[Olivia Lee @olivia_y_lee]

All sensors in. All actuators out: Helix 02 connects every onboard sensor - vision, touch, and proprioception - directly to every actuator through one unified visuomotor neural network.

This unlocks the full dexterity potential of five-fingered hands for intricate manipulation.

[Figure @Figure_robot]

Helix 02’s tactile sensing and palm cameras unlock manipulation tasks beyond pure vision‑based policies

We demonstrate four tasks at the frontier of multi-fingered dexterity

[Figure @Figure_robot]

Introducing Helix 02

It's our most powerful model to date - it's using the whole body to do dishes end-to-end and it's fully autonomous

[Figure @Figure_robot]

Introducing Figure 03

[Figure @Figure_robot]

Figure 03 coming 10/9

[Stanford AI Lab @StanfordAILab]

The Stanford AI Lab community is proud to showcase over 20 research papers at CoRL 2025!

Read about them here : https://t.co/F5Z8Ktnbp8

@corl_conf

[Olivia Lee @olivia_y_lee]

@ajaysub110 Thanks for the question! Our work focused on obtaining explicit rewards in image space for RL, specifically via keypoint/waypoint sequences generated by VLMs, rather than implicit rewards via code or learned reward functions. This helped enable practical real-world execution.

[Olivia Lee @olivia_y_lee]

Enabling robots to improve autonomously via RL will be powerful, and dense shaping rewards can greatly facilitate RL.

Our #IROS2025 paper presents a method leveraging VLMs to derive dense rewards for efficient autonomous RL. ⚡🦾

#Robotics #ReinforcementLearning

🧵1/5

[Olivia Lee @olivia_y_lee]

This work has been a long time coming! Special thanks to my collaborators Annie Xie and @KuanFang, as well as @KarlPertsch @chelseabfinn for their guidance.

📖 Paper: https://t.co/NmKJ1mxMUm
🌐 Website: https://t.co/99L2s3idDV

🧵5/5

[Olivia Lee @olivia_y_lee]

On 4 real-world tasks, KAGI enables success in 30K online fine-tuning steps, improving performance of systems using only sparse rewards.

Furthermore, KAGI shows robustness to a 5x reduction in expert demos, achieving similar performance in 45K online fine-tuning steps. 🚀

🧵4/5