I’m thrilled to share that our paper, “Moving On, Even When You’re Broken: Fail-Active Trajectory Generation via Diffusion Policies Conditioned on Embodiment and Task”, has been accepted to the 2026 IEEE International Conference on Robotics and Automation (ICRA)!

Paper Summary

This work introduces DEFT, a Diffusion-based Embodiment-aware Fail-active Task-conditioned trajectory generator. The motivation is simple: when a robot’s hardware degrades — locked joints, reduced ranges of motion, or restricted velocities — prevailing safety standards demand a fail-freeze response, halting the robot until a human can intervene. We instead push toward fail-active operation, allowing robots to safely keep working under arbitrary actuation failures.

DEFT does this by:

  1. Embodiment Conditioning – A structured vector encodes per-joint position and velocity limits induced by failures, conditioning the diffusion model on the robot’s current “embodiment.”
  2. Constraint Conditioning – A one-hot task encoding selects between unconstrained (e.g., pick-and-place) and constrained (e.g., pushing, wiping) motion primitives in a single unified model.
  3. Start–Goal Inpainting & Output Clamping – Hard-enforces endpoint and joint-limit feasibility throughout the denoising process.

Key Results

Across 4.7k failure conditions and 4.7M trajectories on a 7-DoF Franka Emika Panda:

  • 99.5% success on unconstrained motion vs. 42.4% for RRT.
  • 46.4% success on constrained motion vs. 30.9% for differential IK.
  • Near-parity ID/OOD performance — DEFT generalizes zero-shot to failure conditions never seen during training.
  • Perfect (10/10) real-world task completion on long-horizon drawer manipulation and whiteboard erasing under multi-joint failures, where classical optimization-based methods fail entirely.

Why It Matters

DEFT shows that a single conditioned diffusion policy can absorb the combinatorial explosion of possible failure modes and still produce feasible, multi-primitive plans on the fly — without policy switching, retraining, or hand-engineered recovery rules. This is a concrete step toward robots that don’t shut down when things go wrong, but creatively complete their tasks in space, in factories, and anywhere human rescue isn’t an option.

Huge thanks to my co-authors Yaashia Gautam, Rahul Shetty, Anuj Pasricha, Marco M. Nicotra, and Alessandro Roncone, and to NASA’s Space Technology Graduate Research Opportunity program for supporting this work. See you in Vienna!