Dexterous insertion: testing Play2Perfect’s public policy in Isaac Sim

After box transport, we tested part insertion using the pretrained policy released by Play2Perfect’s authors. TeamZ did not train a new model. The robot is a KUKA iiwa arm with a Sharpa dexterous hand, not G1. In Isaac Sim, it inserts an already grasped L-shaped part into a fixture and withdraws the hand. This is not a physical-robot test.

Insert an already grasped part and withdraw

We hid the auxiliary model showing target placement without changing physics or control. Object and joint trajectories matched across all 172 steps before and after this visual change.

Results

All three initial environments (3/3) and ten environments with another random seed (10/10) satisfied every subgoal and the withdrawal condition in the authors’ evaluator. No early failures or incomplete trials were excluded. The separate recording trial is not included in those counts.

These are small-sample results under the public conditions, not evidence of the same success rate on unknown parts or a physical robot. A 0.5 mm part clearance is different from demonstrating 0.5 mm positioning accuracy. The task clearance is 0.5 mm, but its insertion-position success parameter is 0.01 m. Several seconds of stable rest after withdrawal were not part of this success evaluation.

Execute the pretrained policy, not a replayed motion command

Play2Perfect is a two-stage reinforcement learning method that proceeds from object-manipulation pretraining to assembly tasks. We used its published insertion model without additional teleoperation or training.

The policy has 140-dimensional input including joint and object states. This is not an ACT or camera-based VLA test. Object state comes from the simulator; physical deployment would require a way to measure or estimate it.

From transport to contact-rich tasks

The earlier G1 experiment examined box support and friction. This experiment uses a different robot configuration to reproduce contact-rich part/fixture insertion with a public policy. It does not show G1’s transport capability extending directly to precision assembly.

TeamZ reproduces public methods, builds simulation environments and checks results through logs and video. Contact us about robotics implementation and evaluation.

Primary sources and execution conditions

Play2Perfect paper · Authors’ code, revision ff2fc62873f6c6cd93164123c8358e9ce2d65c9b · Public checkpoints, revision 29f86a6cec2b7bd1aa7ffa8d5031e628a22386e4.

Executed with Isaac Sim 5.1.0.0, Isaac Lab 2.3.2.post1 and NVIDIA L4. The code is MIT-licensed and the model card also lists MIT. We publish our execution results and recordings, not third-party robot models or learned weights.