Horizon-Coupled Obstacle Guidance for Relative-Action Diffusion Policies
TL;DR Steer a frozen diffusion policy around obstacles it never saw in training by pushing each obstacle penalty back onto every relative action that leads into it. On a real Franka, it reaches 105/140 collision-free task successes across seven obstacle scenarios, versus 8/140 for point-wise guidance.
What they do
A diffusion policy trained on obstacle-free demonstrations drives straight into an obstacle placed in its path at deployment. This paper steers a frozen DP3 policy around known static obstacles at inference time, with no new demonstrations and no retraining. The policy outputs relative end-effector actions, so each predicted position is the sum of all earlier increments. An obstacle penalty at a later point can therefore be pushed back onto every increment that led there, which bends the whole action chunk into one detour instead of nudging each step separately.
How it works
- Rollout. At every DDIM denoising step, the clean action estimate is integrated into a predicted end-effector trajectory over the 16-step horizon.
- Obstacle cost. Predicted points, plus 10 interpolated points per segment, are checked against signed distance fields of the measured obstacles. The cost is the squared penetration depth, pushed outward from each obstacle’s center.
- Horizon coupling. The guidance on increment j sums the obstacle responses of every position at or after step j. Point-wise methods such as Lan-o3dp correspond to guiding each step on its own. Only positions are guided; rotation and gripper commands are left to the policy.
- Setup. Franka Panda, one Azure Kinect point cloud, 50 teleoperated demonstrations per task, actions executed at 1 Hz.
Results
- Across 140 real trials (20 per scenario), 105 succeed without collision, against 8 for absolute-action point-wise guidance and 0 for the unguided relative-action policy.
- Pick-and-place with 1-4 obstacles: 18, 16, 12 and 8 of 20. It is the only method that succeeds at all in every layout.
- Pouring, pen insertion and sock-into-drawer with obstacles: 20/20, 16/20 and 15/20.
- Ablation: horizon coupling adds 10-14 successes over point-wise guidance, SDFs add 2-6 over primitive envelopes, and relative actions alone do not help.
Limitations
Obstacles must be static, measured by hand and known in advance. Avoidance covers only the end-effector reference point in the horizontal plane. Guidance strength is tuned per scenario, success falls as free space shrinks, and guidance drops inference from 4.10 Hz to 2.16 Hz.
Figures
Abstract
Real-world robot deployment requires adapting learned policies to unseen obstacles without collecting new demonstrations or retraining. Existing policy-level guidance often applies obstacle corrections independently at individual horizon steps, which may address local obstacle violations but fail to coordinate the action sequence into a complete, task-directed detour. How obstacle responses should be propagated across the action horizon to coordinate complete, task-directed avoidance motion remains underexplored. We introduce an inference-time obstacle-guidance framework for pretrained relative-action diffusion policies. It rolls out cumulative relative actions into a workspace trajectory, evaluates obstacle interactions along the trajectory with an SDF-based obstacle model, and propagates each obstacle response to all contributing action increments during denoising. Real-robot experiments across seven obstacle-constrained scenarios spanning four manipulation tasks show that our method achieves 105/140 (75.0%) collision-free task successes, compared with 8/140 (5.7%) for absolute-action point-wise guidance. Ablations show that the relative-action horizon-coupled formulation provides the larger observed improvement across both obstacle models.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.