Real-World Data for Bi-DexHands

Bi-DexHands provides standardized evaluation for robot learning. Real-world data validates whether simulation performance transfers to physical hardware.

Bi-DexHands at a Glance

Sim
Environment
Multi
Tasks
Standard
Evaluation
Active
Community

Benchmark Profile

Bi-DexHands is a benchmark for bimanual dexterous manipulation using two multi-finger robot hands. Developed at PKU, it provides standardized tasks requiring coordination between two dexterous hands in IsaacGym — a GPU-accelerated simulation environment.

Task Set
20+ bimanual tasks: pass object hand-to-hand, reorient objects cooperatively, open bottles, fold cloth, stack blocks with two hands. Tasks require inter-hand coordination and force distribution.
Observation Space
Joint positions and velocities for both hands (48+ DOF total), object pose, contact forces at fingertips of both hands.
Action Space
Joint position targets for two simulated dexterous hands. 48+ dimensional action space.
Evaluation Protocol
Task-specific success metrics evaluated over parallel GPU environments. Reports success rate, time to completion, and coordination quality metrics.

The Sim-to-Real Gap

IsaacGym GPU parallelism enables massive training throughput but simplified contact between two hands and shared objects. Real bimanual dexterous manipulation involves inter-hand communication delays, asymmetric hand capabilities, and the physical coupling of forces through shared objects that simulation approximates.

Real-World Data Needed

Real bimanual dexterous manipulation recordings with dual-hand force sensing. Object handoff data with authentic timing and force coordination. Deformable object manipulation (cloth, rope) with two hands.

Complementary Claru Datasets

Custom Bimanual Dexterous Collection

Real two-hand manipulation captures the coordination dynamics, timing constraints, and force distribution that simulation simplifies.

Force-Torque Manipulation Dataset

Dual-hand force data during cooperative manipulation provides the coordination signal bimanual policies need.

Bridging the Gap: Technical Analysis

Bi-DexHands pushes dexterous manipulation into the bimanual domain, where coordination between two multi-finger hands creates challenges that single-hand benchmarks do not address. The 48+ dimensional action space (two hands with 24+ DOF each) creates an enormous exploration problem that current RL methods can only handle with GPU-accelerated parallel simulation.

The coordination challenge is the benchmark's defining characteristic. Tasks like passing an object from one hand to another require precise timing of release and catch, force distribution that prevents dropping, and the ability to regrasp if the initial handoff is imperfect. Simulation assumes both hands receive and execute commands simultaneously, but real bimanual systems face communication latency and asynchronous control loops.

Deformable object tasks (cloth folding, rope manipulation) are included in Bi-DexHands but represent the hardest sim-to-real transfer problem. The physics of fabric and rope deformation under multi-finger manipulation involves complex material properties that IsaacGym approximates. Real-world bimanual manipulation of deformable objects provides essential validation data.

The GPU-accelerated training paradigm that Bi-DexHands enables (training across thousands of parallel environments) produces policies that have seen massive simulated experience but zero real-world experience. Real bimanual manipulation data provides the distribution calibration that bridges this simulation-reality gap.

Key Papers

  1. [1]Chen et al.. Towards Human-Level Bimanual Dexterous Manipulation with Reinforcement Learning.” NeurIPS 2022, 2022. Link
  2. [2]Makoviychuk et al.. Isaac Gym: High Performance GPU-Based Physics Simulation for Robot Learning.” NeurIPS 2021, 2021. Link

Frequently Asked Questions

Bi-DexHands is a benchmark for bimanual dexterous manipulation using two multi-finger robot hands. Developed at PKU, it provides standardized tasks requiring coordination between two dexterous hands in IsaacGym — a GPU-accelerated simulation environment.

Real bimanual dexterous manipulation recordings with dual-hand force sensing. Object handoff data with authentic timing and force coordination. Deformable object manipulation (cloth, rope) with two hands.

IsaacGym GPU parallelism enables massive training throughput but simplified contact between two hands and shared objects. Real bimanual dexterous manipulation involves inter-hand communication delays, asymmetric hand capabilities, and the physical coupling of forces through shared objects that simulation approximates.

Yes. Claru coordinates data collection on specific robot platforms and in specific environments to enable direct comparison between simulated and real performance for Bi-DexHands tasks.

Get Real-World Data for Bi-DexHands

Discuss purpose-collected data to validate and improve your Bi-DexHands-trained policies on physical hardware.