01
More success with 8.5% of the samples
On 1,000 out-of-distribution HOT3D test demonstrations, each with an unseen object (203), source grasp (338) or geometry variant (459): 272 MPC-success and 728 MPC-failure.
MPC fails, GNR succeeds
Held-out HOT3D test demonstrations on which DIAL-MPC (3,000 samples) fails and GNR succeeds within S = 256 samples (shown: the first of the 256 samples to pass); both re-simulated from stored controls. Across the 1,000-demo test set: 56.2% vs. 27.2%. Shown at half speed, from just after the grasp is formed; lines trace the fingertips over the last 0.7 s (grey: MPC, blue: GNR).
| Method | Samples | CPU core‑s ↓ | Overall ↑ | MPC‑success ↑ | MPC‑failure ↑ |
|---|---|---|---|---|---|
| IK | 1 | 39.0 | 18.30 | 50.00 | 6.46 |
| DIAL-MPC† | 3,000 | 2689.3 | 27.20 | 100.00 | 0.00 |
| MPPI | 3,000 | 1849.6 | 27.70 | 71.32 | 11.40 |
| iCEM | 3,000 | 3862.1 | 34.70 | 78.68 | 18.27 |
| CMA-ES | 3,000 | 2912.7 | 32.00 | 72.79 | 16.76 |
| GNR | 1 | 48.9 | 18.10 | 48.16 | 6.87 |
| GNR | 64 | 90.9 | 49.60 | 88.97 | 34.89 |
| GNR | 128 | 178.6 | 52.20 | 90.07 | 38.05 |
| GNR | 256 | 354.1 | 56.20 | 92.65 | 42.58 |
† DIAL-MPC generates GNR's training data and defines the MPC-success and MPC-failure splits, so its split success rates are 100% and 0% by construction.
Samples: complete trajectories per demonstration for GNR; samples per iteration × MPC iterations for the sampling-based methods.
