Overview
Enabling dexterous robotic hands to perform complex operations that align with human actions is of great significance. This challenge seeks algorithms that transfer human hand-object manipulation trajectories (e.g., bimanual pen capping) to dexterous robotic hands in simulation, aiming to reproduce physically plausible interactions. Specifically, given a reference motion capture sequence—including the 6D poses of the object(s) and hand(s), as well as the angle of each finger—the algorithm should generate dexterous hand actions that can fulfill the desired object motion in simulator (Issac Gym). Reference motions, estimated via vision-based methods (e.g. MoCap), may contain noise. Participants should develop their algorithms based on the open-source data from OakInk2 and GigaHand. The tasks include manipulations involving
- 1) Single-hand manipulation with a single object.
- 2) Bimanual manipulation with a single object.
- 3) Bimanual manipulation with two separate objects (e.g., pen cap and body).
Evaluations are based on trajectories from the private test set of the OakInk2 and GigaHand dataset, which is currently withheld and will be publicly released around one week before the submission deadline.
Submission deadline for results: October 15, 2025 (11:59PM AoE).
Results will be shared during the HANDS workshop at ICCV 2025
Timeline
| July 7 2025 (Opened) | Website release & registration open |
| July 14 2025 | Training sequences, simulation env & evaluation protocols released |
| September 22 2025 | Evaluation sequences released |
| Model submission Deadline | |
| October 17 2025 | Technical Report Deadline for Challenges (Invited Teams) |
Rules for Participation
This challenge does not strictly separate the training set and evaluation set as in conventional benchmarks. All provided sequences are noisy kinematic data. The core task is to design and train a tracker model that can imitate the human demonstration and learn the dynamics within Isaac Gym.
We distinguish Training Sequences and Evaluation Sequences as follows:
- Training Sequences will be released earlier (July 14) to allow participants to develop and debug their tracker models.
- Evaluation Sequences will be released around one week before the submission deadline. These sequences can also be used for training.
At the final submission deadline (Oct 1 Oct 15), participants are required to submit a set of model weights that can perform rollout within our provided Isaac Gym simulation environment. Training code is not required.
The models will be evaluated by us using a withheld, augmented evaluation set, which are similar but not identical to the released Evaluation Sequences. This evaluation process may take some time.
Methods must outperform the current baseline model ManipTrans to be considered valid.
Dataset and Simulation Env.
The OakInk2 training / evaluation dataset and visualization code are available at huggingface.
The official evaluation environment, baseline implementations, data preparation scripts, and evaluation utilities are available at GitHub.
Contact
hands2025@googlegroups.com
also cclikailinsjtu@gmail.com, rao_fu@brown.edu, siriusyang@sjtu.edu.cn
