Shailesh

Researcher @NUScomputing | MAGIC LAB (@DJiafei)
Working on Foundation models for robotics.
Prev. @IITISM_Dhanbad

[Shailesh @Shailes_h_]

@KhoaVuUmn What about a undergrad 😶‍🌫️

[Jiafei Duan @DJiafei]

Great sharing session at @Lorong_AI !
Good to see many folks in Singapore interested in Embodied AI!

[Shailesh @Shailes_h_]

Your VLA policy needs some RL2 steering 🦾

[Ryohei Sasaki@engineer @rsasaki0109]

RL2-VLA: Adaptive RL Latent Compositional Steering with Test-Time Scaling for Vision-Language-Action Models

Overview of RL2. RL2 relies on base VLA samples to approach the task object, then adaptively applies RL compositional steering to diversify actions toward successful states — when failure is preemptively detected.

[Shailesh @Shailes_h_]

My laptop survived the first run, so I'm calling it an absolute win.

[Jiafei Duan @DJiafei]

Ran the first policy in my new lab at @NUSComputing. Zero-shot, fresh out of the box.

[Shailesh @Shailes_h_]

@Usoppu Should have been me

[Shailesh @Shailes_h_]

@akshen121 I believe you😂

[Shailesh @Shailes_h_]

🦾Action Distribution Visualization PCA reveals that RL² shifts the VLA action distribution closer to the ground-truth action during failure states.

[Shailesh @Shailes_h_]

🎯 Most VLA steering methods break down exactly when they're needed most: under out-of-distribution failures.

We uncover distinct test-time scaling laws 📉📈 for success and failure states, revealing why.

This leads to 𝐑𝐋²-𝐕𝐋𝐀 🚀 — an adaptive RL compositional steering framework that intervenes only when the VLA is likely to fail.

🧵👇

[Shailesh @Shailes_h_]

📊Extensive Evaluation We benchmark RL² on OpenVLA, π₀, and π₀.₅ across simulation (SIMPLER / PolaRiS) and real-world tasks under OOD prompts and task environments. RL² achieves up to +17.5% average improvement over strong steering baselines.

[Shailesh @Shailes_h_]

🛠️RL² Framework We train a lightweight offline RL flow-matching policy conditioned on VLA action expert latents, then compose its flow velocity with the frozen VLA at every flow-matching step. SAFE, a SOTA multitask failure detector, activates steering only upon failure.

[Shailesh @Shailes_h_]

📈Success Scaling Laws The opposite holds during success states. Diversity-inducing steering can unnecessarily perturb already-accurate actions, leading to poor scaling compared to baselines. This motivates adaptive steering only when the base VLA is failing.

[Shailesh @Shailes_h_]

📉Failure Scaling Laws During failure states, action diversity is most beneficial. Diversity-inducing steering methods exhibit the strongest scaling behavior, with RL² compositional steering performing the best by generating actions beyond dominant VLA failure modes.

[Shailesh @Shailes_h_]

🚧 The OOD Challenge

While VLAs excel on in-domain tasks, their performance drops sharply under unseen language instructions and task environments.

💡 Recent work tackles this with inference-time steering, enabling better OOD generalization without costly retraining or additional data collection.

[Shailesh @Shailes_h_]

🔗 Learn more about RL²-VLA:

🌍 Project Page (visualizations & demos): https://t.co/e4091TfZtY
📖 Paper: https://t.co/gw4smsu58J
🛠️ Code: https://t.co/W4p3MSsEn7

[Shailesh @Shailes_h_]

Checkout our latest work RL²-VLA. 🙌
Boosting performance of pretrained VLAs via compositional steering 🔄.

🔗📍 https://t.co/9VhPyaZk6x

[Derek Tan Ming Siang @derektan1995]

🔍 How can we steer pretrained VLAs beyond dominant failure modes in OOD scenarios?

We uncover distinct test-time scaling laws 📈 for steering approaches under success and failure states.

This motivates 𝐑𝐋²-𝐕𝐋𝐀💫, an adaptive RL compositional steering framework.

🧵 👇

[Lucas Ngoo @lucasngoo]

We're growing the @cortexairobot team across San Francisco, Singapore, and Malaysia.

We're hiring Robot Operator Managers, Robot Operators, and Software Engineers to scale real-world robotics data, deployments, and infrastructure for physical AI.

Check out our open roles: https://t.co/UzGimN9E9X

[Shailesh @Shailes_h_]

One year at MARMot Lab, Guillaume Sartoretti's group @NUSingapore, as I wrap up here. Probably the most important arc for me.

Got to work on test-time steering for VLAs, adapting a pretrained policy at inference without touching the weights, and figuring out the hard parts of that.

Thanks to Guillaume for giving me the opportunity, and to @derektan1995 , genuinely one of the best people I've gotten to work closely with. And @sood__shivam , William Teo, Jeric Lew, Srikrishna Iyer for all the great discussions we had along the way. 🙂

[Shailesh @Shailes_h_]

With just 13 demos? Looks impressive!

[LeRobot @LeRobotHF]

VLA-JEPA just dropped in LeRobot 🤖

What makes this model special is that it does not just learn what action to take from a given observation, it also leverages a JEPA world model to learn action-relevant dynamics.

During training, the VLA leverages V-JEPA2 by conditioning its predictor. This clever trick adds a world modeling objective to the training, which also allows pretraining on human videos.
At inference, the world model is dropped entirely, keeping only a standard VLA architecture: Qwen backbone and action head.

The demo here was only fine-tuned on 13 examples, showing great pretraining capability and running in real time on @NVIDIARobotics DGX Spark!

VLA-JEPA is the first world model to be ported to LeRobot, and I feel like it won't be the last 🚀

@Thom_Wolf @ClementDelangue

[Derek Tan Ming Siang @derektan1995]

🚀 Introducing Search-TTA for visual navigation and search (VLN), accepted to #CORL2025 spotlight!✨

🔊 Multimodal: emergent alignment to unseen modalities (text/sound)
🛩️ Online Adaptation: statistical refinement of poor VLM outputs
🦁 AVS-Bench: internet-scale dataset

🌐 Links provided below👇