Sling2Sim2Real teaches robots slingshots with one real-world tug
One-shot elastic ID lets robots learn slingshot policies in simulation, then zero-shot transfer.
Sling2Sim2Real tackles the challenge of elastic object manipulation (EOM), where high-dimensional, nonlinear deformations make policy learning difficult. Traditional approaches require many real-world interactions, which can be destructive (e.g., repeated projectile launches). The new framework from researchers at KAIST uses a one-shot Real2Sim2Real pipeline: a single, non-destructive tug on an elastic band provides enough data to identify elastic parameters via a multi-start system identification method that leverages parameter covariance. This calibrated simulator then trains a manipulation policy for slingshot tasks.
After simulation-based policy learning, the trained policy transfers zero-shot to a real Franka Emika Panda robotic arm. Experiments with elastic bands of varying physical properties and target distances demonstrate that Sling2Sim2Real achieves accurate policy execution and robust generalization with minimal real-world interaction. The work, accepted at IROS 2026, significantly reduces the cost and risk of learning high-strain manipulation tasks, opening the door to more practical elastic object handling in robotics.
- One-shot elastic system identification from a single non-destructive interaction using parameter covariance
- Policy learning in simulation followed by zero-shot Sim2Real transfer to Franka Emika Panda arm
- Robust generalization across different elastic bands and target distances with minimal real-world trials
Why It Matters
Enables safer, cheaper robot learning for high-strain tasks like slingshots, reducing real-world trial needs.