Ever wish a robot could just move to any goal in any environment—avoiding all collisions and reacting in real time? 🚀Excited to share our #CoRL2025 paper, Deep Reactive Policy (DRP), a learning-based motion planner that navigates complex scenes with moving obstacles—directly from point cloud input. w/ Jiahui(Jim) Yang (1/N)

Traditional planners are often too slow for dynamic scenes or rely on full environment knowledge—limiting real-world deployment. We distill their knowledge into a neural network that: • Reacts in real time to moving obstacles (300 Hz) • Operates directly on raw point cloud input (2/N)

DRP is trained entirely in simulation. We first generate 10 million trajectories in simulation to pretrain IMPACT, a transformer-based BC policy. Our diverse data gen enables strong zero-shot generalization to in-the-wild scenes. (3/N)

Pre-training alone is prone to compounding errors, causing minor collisions. We finetune IMPACT via DAgger in simulation, where the teacher leverages Geometric Fabrics to provide locally corrective actions. (4/N)

Even when the goal is blocked, IMPACT is trained to wait safely until the obstruction is cleared—avoiding unnecessary collisions. (5/N)

To further enhance reactivity, we introduce DCP-RMP—a goal proposal module for IMPACT that adjusts goals to prioritize dynamic obstacle avoidance. We call our combined system DRP. (6/N)

DRP outperforms previous SOTA motion planners, such as cuRobo and Neural MP. This work was done @CMU_Robotics with co-lead @Jiahui_Yang6709 as well as @yulongli42, Youssef Khaky, @kenny__shaw, @deepakpathak Website: (7/N)

@Jiahui_Yang6709 @yulongli42 @kenny__shaw @deepakpathak Check out Jiahui’s thread for more details about our method!

This looks awesome

This looks really cool, hope to chat more during CoRL!

Amazing work!

Real-time collision-free navigation is a huge step. Point cloud input to motion planning shows how close robotics is getting to human-like adaptability. #Robotics #Automation

@chris_j_paxton Awesome！

Impressive. Adaptable robots navigating dynamic environments could revolutionize so many fields. What are the potential long-term implications?

Congrats on CoRL acceptance! Learning reactive motion policies directly from point clouds and achieving real-time reactivity in cluttered, dynamic scenes is a major leap for deployable robot navigation. The benchmarks and generalization are especially exciting practical usage.

Oh heck yeah!

Congrats!!! This looks so cool

Thanks Anne!

Huge 🔥🔥 Looks awesome

Thanks Tracy

Looks like you are really confident about the algorithm's safety during the experiments.

Yes we wish this all the time

now do reverse. catch a fly.

人間の代わりに多くのことができるようになりました

sounds promising, but how do you handle unexpected scenarios that weren’t in the training data?

why this guy on video is doing only slow proper movements even without attempt to interrupt the robot? Marketing ads good, quality of testing is below zero. - this robot will not survive

Sounds exciting! Can't wait to see it in action.