We introduce TacFiLM, a lightweight method to incorporate tactile feedback into VLA models. By conditioning visual features with tactile representations via FiLM, robots can feel what they cannot see: contact forces, friction, and compliance. This leads to better performance on contact-rich manipulation tasks.

CDPs add a simple contraction regularizer to diffusion policies, pulling nearby sampling trajectories together to suppress solver and score-matching errors. This yields more robust action generation in offline RL and imitation learning, especially in low-data regimes.

VOCALoco predicts the viability and cost of transport for several pretrained locomotion skills from local heightmaps, then executes the safest, most efficient one. It improves robustness on stairs and transfers zero-shot to real hardware.

Imitation learning policies often fail when robots drift out-of-sample. We propose SCDS, a framework that models policies as contractive dynamical systems, ensuring all rollouts converge regardless of perturbations. The architecture guarantees contractivity by construction, enabling unconstrained optimization. We also provide formal bounds on worst-case and expected loss, and demonstrate strong out-of-sample recovery on manipulation and navigation tasks.

We propose SNDS, a neural imitation learning approach with formal global stability guarantees. By jointly training a neural policy and a Lyapunov candidate, the method ensures safe, predictable behavior beyond the training distribution, overcoming instability and computational issues of prior methods.

Links Paper Project page Slides Missing my first conference Unfortunately, I couldn’t attend my first conference paper presentation at CoRL'23 due to US visa processing delays. I applied for a visa well in advance, but faced significant backlog issues that persisted beyond the conference date. As of then (several months after submission), I was still awaiting visa approval, which had been pending since October 2023. The approach Most imitation learning methods learn to copy expert trajectories, but have no guarantees about what happens when the robot drifts off the demonstrated path. That’s a real problem in practice. Earlier work used stable dynamical systems to at least guarantee convergence back to the goal, but those methods tend to struggle with accuracy on complex, highly nonlinear trajectories, or come with significant computational cost. ...

We propose FedCDI, a framework for causal discovery in federated settings where data cannot leave local sites. Rather than sharing raw samples, clients exchange belief updates over causal graphs, enabling privacy-preserving inference from both shared and heterogeneous interventional data.