- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
The paper proposes R²D-Grasp, a method for post-training a pretrained, wrist-rotation-conditioned grasp diffusion model (DDPM) with binary success rewards from a non-differentiable physics simulator (Isaac Gym). Following TreeGRPO, the reverse diffusion process is split at a fixed step τ into shared denoising prefixes and branched suffixes. A between-prefix advantage is assigned to prefix transitions, and a between-prefix plus λ-weighted within-prefix advantage is assigned to suffix transitions, both optimized with a PPO objective. At inference, an external wrist-proposal module generates candidate wrist orientations, which are ranked by a training-free geometric energy; the diffusion model then generates the wrist translation and finger configuration for the retained orientations. The method is evaluated in the same simulator used for post-training, and on four grasp datasets.
Critical strengths
- Physics-based post-training of grasp diffusion models with a non-differentiable simulator is a relevant and practical problem. Pose-supervised grasp generators do not optimize execution outcomes, and the paper shows a large improvement from post-training (Table 2: 68.0 to 85.3 on UniDexGraspNet, 55.0 to 85.1 on RealDex over the supervised DDPM), together with a smaller distribution shift than full-trajectory DDPO (KL 0.12 vs 0.15) and a lower invalid pre-grasp rate (2.5% vs 10.8%).
- The paper evluates on four datasets, two execution protocols, a zero-shot transfer experiment, and ablations that separately vary the post-training strategy (Table 2), the wrist candidate source (Table 4a), and the yaw selection rule (Table 4b). The shared-prefix diagnostic (Fig. 3) is a useful sanity check that late denoising steps still change the physical outcome.
Critical weaknesses
- Novelty over TreeGRPO is unclear, TreeGRPO is not a baseline, and the gain over DDPO is marginal. The sampler is TreeGRPO with a single branch point (M = 2 prefixes, K = 4 suffixes). TreeGRPO already normalizes rewards within a group and propagates them up the tree, so the global/local split in Eq. 19 and the cancellation in Eq. 6 follow directly from it. The only differences are separate normalization of the two parts with a weight λ, training the prefix step and uniform instead of probability-weighted averaging. The introduction's claim that TreeGRPO applies one combined advantage across the whole trajectory contradicts TreeGRPO's own description of step-specific advantages. Despite this, TreeGRPO is not compared against, and λ and the branch step τ are never reported. This matters because the measured effect is small: in the only controlled comparison (Table 2), R²D-Grasp beats full-trajectory DDPO by 1.7 and 2.4 points with no error bars or seed count, and Fig. 3b shows the gain over DDPO ranging from roughly zero to about 3 points depending on the branch ratio.
- Most of the headline gain comes from components that are not the paper's contribution, and the evaluation setting is narrow. Table 4a shows the UniDexGrasp wrist predictor adds 17.4 and 21.6 points and the geometric ranking another 4.0 and 6.0, while the proposed credit assignment adds 1.7 and 2.4 (Table 2); the large jump over the supervised model is already obtained by plain DDPO. The geometric energy is only needed because the wrist orientation is excluded from the diffusion model, and the given justification (one rotation supports several translations) applies equally to a model that generates the rotation as well. Table 1a thus compares a multi-component pipeline against single end-to-end generators. The tabletop setting fixes palm-down orientation and wrist position, so wrist selection reduces to picking 4 of 16 yaw angles for a single settled object with an open-loop approach, while the dynamic baselines run closed-loop pipelines; gradient-based wrist steering is not compared. All results are in Isaac Gym with no sim-to-sim or real-robot transfer, and no table reports standard deviations. The premise that early steps fix the grasp mode and late steps refine contacts is not directly tested; Fig. 3 only shows that late steps can still change the outcome.
Questions for authors
- How does TreeGRPO (depth-1 tree with standard group normalization and leaf-to-root aggregation) perform under the same rollout budget as Table 2? This is the most important missing comparison.
- How many seeds were used for Tables 1 to 4, and what is the standard deviation? Fig. 3b reports five runs with error bars; the tables report none.
- Which wrist candidate source do the Table 1a baselines (GraspTTA, DexGraspAnything) use? If they do not receive the wrist predictor and geometric ranking, Table 1a conflates backbone, wrist handling, and post-training.
- Why is the wrist orientation not diffused jointly with the translation and finger configuration? Was this tried?
- Why M = 2 shared prefixes? With M = 2 the global advantage carries no magnitude information. Were larger M evaluated?
- What values of λ and τ were used, and how sensitive are the results to them?