Sean Woo · 한국어 원문
Key takeaways
- Diffusion Policy is an imitation-learning method that repeatedly refines a noisy sequence of actions, using camera observations as a condition, to produce robot movements.
- When a task has several valid solutions, such as going left or right, it can represent these alternatives and generate a coherent route instead of simply averaging the demonstrations.
- It predicts a sequence of future actions, executes only the first part, then observes the scene again and replans.
- The extended paper reports a 46.9% average improvement over the strongest earlier methods across 15 tasks from four manipulation benchmarks, covering simulation and real-world settings.
- These results apply to the tested conditions. They do not establish a general-purpose robot that automatically handles unfamiliar objects, lighting, camera positions or safety requirements.
When there is more than one way to reach a cup
Imagine a robot reaching for a cup behind a box. A person could demonstrate going around the left side or the right side. Both demonstrations succeed. But if a learner simply averages their coordinates, the resulting movement could point straight into the box. When there are several correct answers, the average action can be a failing action. This scene is an illustrative example, not a reported experiment.
Diffusion Policy addresses the problem of collapsing different possibilities into a single average. Instead of outputting the robot action in one step, it starts with a nearly random sequence of future actions and gradually refines it using camera images and robot-state information. It can represent several peaks in the distribution of possible actions while generating a consistent choice, such as a left-hand or right-hand route, for a particular execution. The official project page explains this ability to model multiple action modes.
A diffusion model generates actions instead of pictures
An image-generating diffusion model starts with noise and repeatedly recovers shapes and details. Diffusion Policy uses a related idea, but its output is a sequence of robot actions, such as positions, rotations and gripper commands, rather than a picture. Camera observations and robot-state information condition the calculation. At each step, a neural network predicts how to reduce the noise in the candidate action sequence.
In the paper, the sequence begins as Gaussian noise and is refined through K denoising iterations. Here, denoising does not mean filtering electrical noise out of a motor signal. It means repeatedly adjusting a computational candidate toward actions consistent with the demonstrations. The researchers compared a temporal one-dimensional CNN with a Transformer. They describe the CNN as a stable default across many tasks; the Transformer can help with rapid changes in action, but is more sensitive to hyperparameters. See the extended paper for the architectures and their evaluation conditions.
Execute a little, then look again
The robot does not blindly execute the entire predicted sequence. It uses recent observations to predict a group of future actions, carries out only the beginning, and then reads the cameras and robot state again before generating a new sequence. This is called receding-horizon control. It resembles navigation software that plans a longer journey but updates its route as the current position changes. The analogy describes replanning, not how the robot learns its actions.
The approach balances smoother, coordinated sequences with responsiveness to change. In the paper’s real-world Push-T setup, the policy generated commands at 10 Hz, and execution used linear interpolation at 125 Hz. Separate tests included blocking the camera with a hand for three seconds and moving the T-shaped block. The robot replanned in these demonstrations, but they do not establish safety under every possible disturbance.
What do 46.9% and 95% actually mean?
The 2024 extended paper evaluates 15 tasks across four robot-manipulation benchmarks. The evaluation includes simulated and physical environments, actions with two to six degrees of freedom, single-task and multi-task settings, and rigid objects as well as liquids. The authors report a 46.9% average improvement over the strongest earlier methods in their comparison. This is an average improvement calculated from the paper’s task scores; it does not mean a 46.9-percentage-point increase in the success rate of every robot task.
In the real-world Push-T task, the robot had to push a T-shaped block onto a target region and then move its end effector to an end zone. The end-to-end Diffusion Policy achieved a 95% success rate and a mean intersection-over-union, or IoU, of 0.80. IoU measures how much the block overlaps the target relative to their combined area. In the same table, position-controlled LSTM-GMM achieved 20% success and IBC achieved 0%. Success was defined using final IoU above the minimum IoU observed in the human demonstration data. The table also reports 100% success and 0.84 mean IoU for human demonstrations. These are results for that task and definition, not factory-wide availability or a safety-reliability figure. See Table 6 and the real-world evaluation.
Outside the demonstrations, validation starts again
Diffusion Policy learns by imitation, so suitable demonstration data is its starting point. Large changes in objects, lighting, camera placement or friction can lead it outside the conditions represented in training. Repeated denoising also takes computation time, so the control cycle and available hardware must be considered together. The paper describes using DDIM to accelerate inference with fewer denoising steps than in training. Worst-case latency still needs to be measured on the intended hardware.
Representing several action possibilities does not ensure that every generated action is safe. Collision constraints, force limits, emergency stopping and work-area monitoring need their own design. The official repository also specifies concrete hardware for reproducing real-world Push-T, including a UR5-series robot, two RealSense D415 cameras and a SpaceMouse for teleoperation. Being able to run the code does not establish immediate transfer to a different robot.
Evaluating a robot therefore requires more specific questions than whether one demonstration video ends successfully. How many starting positions and object types were tested? How many repetitions were performed under each condition? Did failures involve collisions or human intervention? What were the average and worst-case delays from sensing to motor commands? Diffusion Policy shows the value of representing robot behavior as several possible continuous action sequences. Readiness for an actual deployment still depends on data and safety testing for the intended task.
Sources
- Official Diffusion Policy project
- arXiv record — first submitted March 7, 2023.
- Extended paper, version 5 — March 14, 2024.
- Official code and reproduction requirements
About Sean Woo
I have spent more than 15 years working on robotics technology and business direction. I use publicly available technical documents, research papers and company announcements to explain changes in robotics and AI. The interpretations presented by this publication do not represent the official views of any company or institution.