Model Card for GraspGen (IB-Robot)

GraspGen (Generative Grasp Sampler with Diffusion Denoiser + Discriminator) for point-cloud-based robotic grasp generation, packaged for the IB-Robot framework with two deployments sharing one contract.

Deployments

deployment	backend	artifacts	notes
ascend_310p	Ascend ACL (Ascend310P1)	8 OM modules	board-side pipeline (generator/discriminator split graph, device links)
torch_cuda	PyTorch CUDA	— (weights in assets/)	host-side debugging and evaluation

Both deployments share the same tensor_model/graspgen/generate_grasps contract (observation.object_points [-1,3] -> grasp.poses [-1,4,4] + grasp.confidence [-1]).

Weights provenance

• assets/generator_checkpoint.pth = NVlabs/GraspGen checkpoints/graspgen_robotiq_2f_140_gen.pth (sha256 fe8497108e39d8fc50be06cd7df22a2f680e0495e713d24c3104616f291e00dd)
• assets/discriminator_checkpoint.pth = graspgen_robotiq_2f_140_dis.pth
• The 310P OM modules were converted from the same checkpoint — both deployments are weight-identical (sha pinned in assets/adapter.json).

Repository Structure

• inference_manifest.json — deployment routing (schema v3)
• assets/adapter.json — algorithm contract, runtime parameters, checkpoint digests
• assets/graspgen_config.yml — upstream omegaconf config (robotiq_2f_140)
• assets/generator_checkpoint.pth / assets/discriminator_checkpoint.pth — Torch weights
• artifacts/ascend/ascend_310p/*.om — compiled 8-role pipeline

Usage

Host (CUDA, requires the grasp_gen package with pointnet2_ops):

from inference_manifest import load_inference_manifest
validated = load_inference_manifest("models/grasp", "torch_cuda")
# driven via manipulation_service.graspgen_wrapper (inference_backend="local_cuda")

Board (Ascend 310P): select the ascend_310p deployment through the IB-Robot unified inference runtime.

License

Code and packaging: Apache-2.0. GraspGen model weights follow the NVlabs/GraspGen upstream license (NVIDIA Source Code License) — check compatibility before redistribution.