DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning
AuthorsMinho Park, Kinam Kim, Donghu Kim, Byungkun Lee, Dongyoon Hwang, Yongjae Shin, Junha Hyung, Hojoon Lee, Jaegul Choo
AffiliationsKAIST · Holiday Robotics
DiscoDemo trains robots to find varied successful ways to complete tasks, then turns those behaviors into better imitation-learning data.
Key results
Average across four tasks, compared with 40% for the strongest baseline.
Average across four tasks, compared with 39% for the strongest baseline.
Dataset size used to train and compare downstream policies.
Evaluation covered object placement, stacking, and two peg-insertion tasks.
What the paper found
DiscoDemo treats reinforcement learning as a way to generate training data, not as the robot’s final controller. Its privileged generator combines a parallel reverse curriculum with latent-conditioned skills and a success-gated METRA-style diversity reward, encouraging different ways to complete a task without rewarding diverse failures. Safety shaping and filtering then produce a static set of successful demonstrations for a downstream visuomotor policy. Across 4 manipulation tasks in NVIDIA Isaac Sim and on a real robot, the method trains the same DROID-pretrained π0.5 vision-language-action model using 3000 generated demonstrations per data source. Average success reached 62% in simulation, versus 40% for the strongest baseline, and 78% on the real robot, versus 39%. The results suggest that broader coverage of successful behaviors makes imitation policies more robust, while excessive diversity can undermine task success; DiscoDemo’s gains also complement corrective data collection such as DAgger and DART.
Original abstract
Behavior cloning for robotic manipulation is highly sensitive to the coverage of its training data, yet collecting diverse successful demonstrations remains costly. We study reinforcement learning (RL) not as a deployable controller, but as a data generator for downstream visuomotor imitation learning. Our key premise is that multiple successful solution modes cover a broader range of successful trajectories than repeatedly collecting a single narrow solution. We introduce DiscoDemo, which trains a privileged latent-conditioned RL generator with a parallel reverse curriculum and a success-gated diversity objective to discover diverse solutions while preserving task success. The trained generator produces multiple successful closed-loop behaviors, which are safety-filtered and exported as a static dataset for downstream imitation learning. Across four manipulation tasks, DiscoDemo produces broader successful-state coverage and improves downstream visuomotor policies in both simulation and the real world. Averaged across tasks, downstream success increases to 62% from 40% for the strongest baseline in simulation, and to 78% from 39% on the real robot, respectively. These results suggest that successful solution diversity improves imitation learning for a pretrained flow-matching VLA and remains complementary to corrective data collection. Project page: https://davian-robotics.github.io/DiscoDemo
Read the original paperMore in Robotics
Browse all 59 papers →HoloBoxing: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training
Xihan Ma, Bo Zhang, Boxiong Li, Yuhan Bao, Ziyue Liu, Yucheng Wang
A pretrained humanoid controller helps a compact planner learn simulated boxing in about seven minutes of additional single-GPU training.
HyperDrone: Simulation-to-Reality Transfer at Millions of Frames per Second
Jonas Eschmann, Giuseppe Loianno
HyperDrone aims to make realistic drone simulation fast enough to train vision-based flight policies at massive scale, then transfer them to a real quadrotor.
Blind Precise Assembly: Proprioceptive Sim-to-Real Peg-in-Hole for Diverse Parts
Antonia Bronars, Younghyo Park, Kushal Kedia, Pulkit Agrawal
A robot learns to insert, mesh, and thread parts by feeling contact through its own joints, without vision or tactile sensors.