TL;DR
DexSinGrasp introduces a unified reinforcement learning framework that enables a multi-fingered dexterous hand to singulate (separate) and grasp target objects in dense clutter. By combining a unified reward function, a dense-to-random curriculum learning strategy, and teacher-student distillation into a vision-based policy, the robot learns emergent finger-level clearing behaviors—substantially outperforming multi-stage and grasp-only baselines in both speed and success rate.
What Problem Does This Paper Address?
Grasping a target object packed tightly among obstacles is a classic robotic manipulation challenge. While traditional two-fingered parallel jaw grippers must push obstacles out of the way using broad arm movements—often requiring tedious, multi-step push-then-grasp routines—dexterous anthropomorphic hands can theoretically flick, nudge, and clear obstacles using only their fingers.
However, controlling high degree-of-freedom (DoF) dexterous hands in dense clutter is notoriously difficult. Prior approaches either ignore singulation (hoping to find collision-free grasp paths, which fails in tight clutter) or split singulation and grasping into disjoint, sequential policies that switch via hard-coded thresholds. DexSinGrasp tackles this by training an end-to-end policy that seamlessly unifies dexterous singulation and grasping into a single continuous action space.
Methods
The authors formulate the task as reinforcement learning optimized via Proximal Policy Optimization (PPO), followed by vision-based distillation:
- Unified Piece-Wise Reward Formulation: Instead of separating the problem into discrete stages, the policy optimizes a piece-wise reward function. Before contact, it rewards approaching the target and penalizes distance, while simultaneously rewarding object singulation ($r^S$, incentivizing pushing clutter away). Once hand-object contact thresholds are met ($d^P < 0.06\,\text{m}$ and $d^J < 0.2\,\text{m}$), it adds lifting and goal-reaching rewards while retaining the singulation bonus to keep surrounding clutter clear.
- Clutter Arrangement Curriculum Learning: Learning in heavy clutter from scratch is prone to local optima. The authors progressively train privileged state-based teacher policies across increasing task difficulty: starting from a single isolated object $\rightarrow$ dense clutter (D-4 $\rightarrow$ D-6 $\rightarrow$ D-8 blocks) $\rightarrow$ random clutter distributions (R-4 $\rightarrow$ R-6 $\rightarrow$ R-8).
- Teacher-Student Policy Distillation: Because privileged state information (such as ground-truth obstacle positions and singulation distances) is unavailable in real-world deployment, the authors distill the teacher policies into a vision-based student policy via Behavior Cloning. The student receives robot proprioception and point clouds processed by a pre-trained, frozen point cloud encoder (from UniGraspTransformer).
Datasets
- Clutter Benchmark (Simulation): Built in Isaac Gym, featuring tabletop environments where a central target block is hemmed in by 4, 6, or 8 surrounding obstacle blocks of varying shapes ($1\times1$, $1\times2$, $1\times3$). Tasks are categorized into Dense (D-4, D-6, D-8) tightly packed clusters and Random (R-4, R-6, R-8) clutter arrangements.
- Real-World Setup: Evaluated on a 6-DoF uFactory xArm6 robot arm equipped with a 16-DoF LEAP Hand and two RealSense D435 RGB-D cameras fused via Iterative Closest Point (ICP) to generate a 1,024-point filtered point cloud at 20 Hz.
Key Insights and Ideas
- Emergent Dexterous Micro-Motions: Rather than relying on clumsy arm pushes, the unified policy discovers subtle, highly dynamic finger-level singulation strategies: finger flickering, palm rubbing, and finger-palm vibration to jostle, nudge, and destabilize surrounding obstacles just enough to slip fingers in for a grasp.
- Curriculum Direction is Crucial: The sequence of task difficulty matters significantly. Training in a dense-to-random order (mastering tight geometric constraints before adapting to varied spatial clutter) outperformed random-to-dense. In contrast, training from scratch on the hardest tasks failed catastrophically (achieving 0% success on the R-8 task).
- Unified Synergies Beat Two-Stage Pipelines: Breaking the task into explicit "singulation" and "grasping" sub-policies (with distance threshold triggers) creates rigid failure modes and inflates execution time. A unified reward allows the hand to singulate and grasp simultaneously.
Key Results
- Dense Clutter Performance: In simulated dense clutter (D-4 to D-8), the teacher policy achieves a 98% average Success Rate (SR) in 114 steps (AS). The distilled vision-based student policy achieves 89% SR in 114 steps. In comparison:
- GraspReward-Only Baseline: 39% SR / 185 AS (drops to 10% on D-8).
- Multi-Stage Singulation Baseline: 72% SR / 183 AS.
- Random Clutter Performance: On random arrangements (R-4 to R-8), the teacher policy reaches 96% SR (88 AS), and the vision student achieves 88% SR (84 AS), substantially outperforming the Multi-Stage baseline (79% SR / 133 AS) with ~37% fewer simulation steps.
- Real-World Deployment: The distilled student policy zero-shot transfers to the physical LEAP Hand setup, successfully singulating and picking target blocks across all D-n and R-n scenarios.
Limitations and Future Work
- Geometric Simplicity: Obstacles and targets are currently restricted to primitive block shapes ($1\times1$, $1\times2$, and $1\times3$). Evaluating on diverse, daily household geometries and varied friction/mass distributions remains an open challenge.
- Static Clutter Assumptions: While surrounding objects move when touched, the initial clutter configurations are static; dynamic environments (e.g., shifting piles or moving obstacles) were not explored.
- Sensing Occlusions: Real-world grasping in dense clutter suffers from severe optical occlusions, meaning the distilled policy occasionally relies on imprecise point cloud centroids when deep inside clutter. Incorporating tactile feedback into the student policy is a natural next step.
Summary updated