TL;DR
This paper introduces a novel inference-time guidance framework that allows pre-trained robot "Diffusion Policies" to navigate around new obstacles without any retraining. By coupling obstacle responses across the entire predicted action sequence—rather than correcting each step in isolation—the method enables robots to perform complex detours while maintaining their original task goals, achieving a 75% success rate compared to just 5.7% for standard approaches.
What Problem Does This Paper Address?
Visuomotor policies based on Diffusion Policy (DP) have become a gold standard for robot manipulation because they can learn complex, multi-modal behaviors from human demonstrations. However, they have a major weakness: they are "brittle" to changes in the environment. If you train a robot to pick up a cup in an empty space and then place a bottle in its path, the robot will likely collide with the bottle because the obstacle wasn't in its training data.
Retraining a policy for every possible obstacle configuration is prohibitively expensive. Researchers have tried "inference-time guidance"—nudging the robot's predicted actions away from obstacles during the generation process. The problem is that most existing methods apply these nudges point-wise. They look at each step in a predicted action "chunk" independently. While this might stop a robot from hitting a wall at one specific moment, it often fails to coordinate the entire sequence into a smooth, goal-directed detour. The robot might avoid the obstacle but then "forget" where it was going or end up in an unreachable configuration.
Methods
The authors propose Horizon-Coupled Obstacle Guidance, specifically designed for policies that predict relative actions (incremental displacements rather than absolute coordinates).
The core technical innovation lies in how the guidance is calculated:
- Trajectory Rollout: During the denoising process (where the policy generates an action chunk), the framework "rolls out" the predicted relative increments into a full workspace trajectory.
- SDF Evaluation: It evaluates this trajectory against a Signed Distance Field (SDF)—a mathematical map that tells the robot how far it is from an obstacle and in which direction "safety" lies.
- Horizon Coupling: Because the robot uses relative actions, the position at step 10 depends on the sum of all increments from steps 1 through 9. The authors use the Jacobian (a matrix of partial derivatives) of this summation to propagate obstacle responses backward through the entire horizon.
Essentially, if the robot predicts it will hit something at step 10, the framework doesn't just nudge step 10; it adjusts all the preceding steps to ensure the entire path is a coherent detour.
Datasets
The researchers didn't use a standard static dataset but rather collected their own expert demonstrations to train the base policies:
- Demonstrations: 50 obstacle-free teleoperated demonstrations for each of the four tasks.
- Evaluation Scenarios: Seven real-robot scenarios with varying obstacle layouts (single blocks, narrow passages, and multi-obstacle constraints).
- Tasks: Pick-and-place, Pouring (requires end-effector roll), Pen insertion (requires pitch adjustment), and a sequential "Sock + Drawer" task.
Key Insights and Ideas
- Relative Actions are a Feature, not a Bug: While absolute actions are common, relative actions naturally create a mathematical dependency across the time horizon. The authors realized that the simple act of summing increments ($p_i = p_0 + \sum \Delta a_j$) provides the perfect structure for "coupling" the guidance.
- Global Coordination vs. Local Correction: Point-wise guidance is like a driver only looking at the front bumper; horizon-coupled guidance is like a driver looking at the whole road ahead. This global view is what allows the robot to maintain task progress while avoiding collisions.
- Zero-Shot Adaptation: The most powerful aspect is that the base policy is frozen. You can take a policy trained in a clean lab and deploy it in a cluttered kitchen immediately, provided you have a basic geometric model (SDF) of the new obstacles.
Key Results
The performance jump was dramatic. In real-robot experiments across four manipulation tasks:
- Success Rate: The proposed method achieved 105/140 (75.0%) collision-free task successes.
- Baseline Comparison: The standard absolute-action point-wise guidance (representing the current state-of-the-art for policy guidance) managed only 8/140 (5.7%).
- Behavioral Consistency: As shown in the trajectory plots below, the coupled guidance (bottom row) produces consistent, smooth detours, whereas the baseline (top row) often results in trajectories that terminate prematurely or wander aimlessly.
Limitations and Future Work
- Static Environments: The current framework assumes obstacles are static and their positions are known (to build the SDF). It cannot yet handle moving obstacles or dynamic environments in real-time.
- Manual Modeling: Obstacle geometry was measured manually for these experiments. Integrating automated scene reconstruction (e.g., using NeRFs or real-time point cloud processing) is a logical next step.
- Computational Overhead: Adding guidance during the denoising process increases inference latency (from ~4 Hz down to ~2 Hz in their tests). While still usable for many tasks, further optimization is needed for high-speed applications.
Summary updated