Satyajeet Das

Steering Multirobot Behavior via Closed-Loop Affine Activation Editing

Satyajeet Das, Darren Chiu, Shashank Hegde, Gaurav S. Sukhatme

arXiv

CLAE enables closed-loop behavior steering of frozen robot policies through small, state-dependent edits to internal activations at inference time. It provides a lightweight way to adapt pretrained policies to new behavioral objectives, including injecting behaviors beyond the base policy’s original objective, without modifying the base policy weights, demonstrated across individual and coordinated multirobot behaviors.

Data-Efficient Adaptation of a Driving VLA to Class 8 Trucks

arXiv coming soon

We adapt a pretrained driving VLA to Class 8 trucks using a few hundred real-world scenarios. Our adapt-then-steer approach combines targeted fine-tuning with lightweight inference-time steering to improve trajectory predictions in long-tail truck-driving scenarios.

Latent Activation Editing: Inference-Time Refinement of Learned Policies for Safer Multirobot Navigation

Satyajeet Das, Darren Chiu, Zhehui Huang, Lars Lindemann, Gaurav S. Sukhatme

ICRA 2026

LAE steers pretrained robot policies at inference time by editing internal activations instead of retraining the policy. A latent collision world model predicts pre-collision activations and uses them to replace unsafe activations online, reducing collisions by nearly 90% in multi-quadrotor navigation while preserving task completion.

WARPD: World model Assisted Reactive Policy Diffusion

Shashank Hegde, Satyajeet Das, Gautam Salhotra, Gaurav S. Sukhatme

arXiv

WARPD shifts diffusion-based robot learning from generating open-loop trajectories to generating reactive closed-loop policies. Using latent diffusion and a co-trained world model, it enables longer action horizons, robustness to perturbations, and substantially lower inference cost while maintaining strong task performance.

Latent Weight Diffusion: Generating Reactive Policies instead of Trajectories

Shashank Hegde, Satyajeet Das, Gautam Salhotra, Gaurav S. Sukhatme

NeurIPS 2025 Embodied World Models for Decision Making Workshop, RSS 2025 Resource constrained robotics workshop

Latent Weight Diffusion (LWD) introduced the shift from trajectory generation to policy generation, where a diffusion model learns a latent trajectory space and a hypernetwork decoder transforms it into policy weights, yielding reactive closed-loop policies. This approach delivers long-horizon robustness and achieves ~45x lower inference compute while maintaining SOTA performance across manipulation and locomotion tasks.

Real-Time Perception Based Control Barrier Functions for Efficient Robotic Navigation Using Depth Camera

Satyajeet Das, Yifan Xue, Haoming Li, Nadia Figueroa

SII 2026

RNBF brings NeRF (Neural Radiance Fields)-inspired 3D scene reconstruction directly into the robot's control loop by learning continuous, differentiable neural signed distance fields (SDFs) online from noisy RGB-D input. Unlike prior methods that are offline or require pretraining, RNBF updates its neural SDFs in real time (5–15 Hz) and feeds them into standard Control Barrier Function (CBF-QP) controllers, enabling safer navigation in unknown environments using only a depth camera.

On the Feasibility of EEG-based Motor Intention Detection for Real-Time Robot Assistive Control

Ho Jin Choi ^*, Satyajeet Das ^*, Shaoting Peng ^*, Ruzena Bajcsy, Nadia Figueroa

ICRA 2024

* Equal contribution: Ho Jin Choi, Satyajeet Das, and Shaoting Peng.

This work demonstrates the feasibility of using EEG signals for real-time motor intention decoding to control assistive robots. Using Riemannian geometry features from EEG covariance matrices and a lightweight SVM classifier, the system achieves ~70% online accuracy in robot-in-the-loop experiments for left/right arm intention prediction. These results establish a practical pathway toward non-invasive brain–robot interfaces for assistive robotics.

Learning Robust Output Control Barrier Functions from Safe Expert Demonstrations

Lars Lindemann, Alexander Robey, Lejun Jiang, Satyajeet Das, Stephen Tu, Nikolai Matni

IEEE Open Journal of Control Systems

Designed a ResNet-based CNN to map dashboard camera images to a perception representation, supporting a perception-based robust control barrier function implemented with a two-layer DNN.