Derek Tan Ming Siang

PhD Candidate @ NUS 🇸🇬, Robotics VLA/VLN | Visiting Researcher @UofTRobotics 🇨🇦

[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.

🧵 👇

[Derek Tan Ming Siang @derektan1995]

@Shailes_h_ Very cool!

[Derek Tan Ming Siang @derektan1995]

@NateTechFuture We report real-robot evaluations over 4 tasks. Do check out our paper for more details: https://t.co/aRZBAmFXIY

[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.

🧵 👇

[Derek Tan Ming Siang @derektan1995]

Huge thanks to my amazing co-lead and collaborators @Shailes_h_ @kkrish_i @williamteo @AveryJuuu0213 @qiaogu1997 & Prof Guillaume for all the hard work and guidance throughout this project!

[Derek Tan Ming Siang @derektan1995]

📊(6) 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.

[Derek Tan Ming Siang @derektan1995]

🦾(8) Action Distribution Visualization

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

[Derek Tan Ming Siang @derektan1995]

🛠️(5) 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.

[Derek Tan Ming Siang @derektan1995]

📈(4) 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.

[Derek Tan Ming Siang @derektan1995]

📉(3) 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.

[Derek Tan Ming Siang @derektan1995]

🚧(2) The OOD Challenge

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

Recent works tackle this through inference-time steering, improving OOD generalization without costly retraining or data collection.

[Derek Tan Ming Siang @derektan1995]

📖(1) Check out more visualizations on our website. We also made it easy for you to try our code:

🌐 Website: https://t.co/aRZBAmFXIY
📄 Paper: https://t.co/pHvC6RQmqX
💻 Code: https://t.co/gSiHvPdsSJ

[Derek Tan Ming Siang @derektan1995]

@Shailes_h_ @NUSingapore Cheers to the best RA we can ever ask for, and best of luck with your PhD journey ahead!

[Derek Tan Ming Siang @derektan1995]

@harine_ravi Hi Harine, I'll be at CoRL too, presenting my work on online adaptation for VLNs! Would love to chat :)

[Derek Tan Ming Siang @derektan1995]

For more info:
https://t.co/iuErT1NtN5

[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👇

[Derek Tan Ming Siang @derektan1995]

📍I'm excited to be at Seoul for #CORL2025 with my co-authors for our spotlight presentation on Search-TTA this coming Monday! 🇰🇷

🔊 Multimodality: emergent alignment to unseen input modalities
🛩️ Online Adaptation: refine poor VLM predictions
🦁 AVS-Bench: internet-scale data

[Derek Tan Ming Siang @derektan1995]

We will be presenting in Spotlight session 4 and Poster session 2 (3:30-6:00pm, 29 Sept). Looking forward to meeting all of you there!

[Derek Tan Ming Siang @derektan1995]

🙏 Special thanks to my supervisors Prof Florian Shkurti and Prof Guillaume Sartoretti, who made the research visit to Toronto possible and for their solid guidance. Also to my co-authors and friends from NUS, Toronto, and ST Engineering who made this journey so enjoyable!

[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👇

[Derek Tan Ming Siang @derektan1995]

💻 Framework Flexibility:
We design Search-TTA to be flexible to different input modalities, planners, and environments. Applications include multimodal object detection, environmental monitoring, household search etc. Customize our code for your use case:
https://t.co/Npkv3LUvRm