DAPPT: Domain-Aware Policies from Procedural Trajectories for Sim-to-Real Dexterous Tool Use

TL;DR Train a dexterous hand-arm policy in rigid-body-only simulation to track procedurally generated tool trajectories, with the trajectory parameters as a policy input that can be tuned per task. Picking the best parameters raises task metrics 1.5-3.8x in a particle simulator, and the policy transfers zero-shot to a real robot on six tasks such as sand scooping and play-dough rolling.
What they do
Tool use with sand, dough or liquids is slow to simulate and hard to transfer, which rules out large-scale RL. DAPPT never simulates the non-rigid material. It simulates only the rigid tool and trains a policy to track procedurally generated tool trajectories, such as a scoop entering sand at a given angle, depth and speed. The trajectory parameters (“domain parameters”) are a policy input, so at test time one can search for the setting that gives the best task result.
How it works
- Procedural trajectories. Each task defines a parameterized 6D pose or joint trajectory of the tool, plus a contact trajectory saying which hand links touch the tool and where.
- Teacher. PPO in IsaacLab trains a state-based policy for a Franka Research 3 arm with a 12-DoF XHand1 hand, with randomized domain parameters, contact and pose-tracking rewards, a decaying behavior-cloning term from 10-20 simulated demos, and domain randomization. Arm damping gains are calibrated to real open-loop trajectories.
- Student. DAgger distills the teacher into a PointNet++ point-cloud policy that also takes the domain parameters.
- Parameter selection. Random search in Genesis, which simulates sand and dough with MPM, picks the best parameters per task.
Results
- Best vs. average parameters improves task metrics 1.5-3.8x in Genesis (scooped sand: 3.64 g to 13.9 g).
- DAPPT beats SimToolReal on all six tasks; that baseline, even with ground-truth object poses, scores near zero on scraping, rod-rolling, dispenser and spray.
- Training only on the best parameters instead of all domains drops the task metric 49-98% on the three tasks tested.
- Real robot, zero-shot, 20 trials per task: task success 14/20 scooping, 11/20 rod-rolling, 19/20 scraping, 19/20 mouse-wheel, 14/20 dispenser, 10/20 spray (no water emitted). Real scooping gets 7.7 g with the best parameters vs 2.9 g with mean parameters.
Limitations
Policies generalize only to a narrow range of object poses and assume a known object. Domain parameters and trajectory functions are hand-designed for each task.
Figures




Abstract
Sim-to-real RL has achieved impressive results in dexterous manipulation but remains limited to tasks with accurate rigid-body simulation. However, these methods cannot be applied to skills that involve deformable and particles, with which the simulation speed will become too slow for large-scale RL learning. We present DAPPT, a sim-to-real framework for end-to-end dexterous tool-use manipulation skills, e.g., sand scooping, play-dough rod-rolling, without simulating non-rigid materials. Our method learns a domain-aware RL policy by tracking procedural trajectories parameterized by domain parameters, and distills it into an end-to-end domain-aware sensorimotor policy. Furthermore, explicit domain represen-tation enables fast selection of domain parameters that can maximize downstream task performance. We evaluate DAPPT end-to-end on six dexterous manipulation tasks, including sand scooping, play-dough rod-rolling, sand scraping, mouse-wheel scrolling, dispenser pressing, and spray triggering. Our method achieves 1.5-3.8x in task metrics through the best domain parameter and transfers zero-shot to the real world.
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.
