UNOGrasp: Unified Obstruction Reasoning

• UNOGrasp is a vision-language model that formulates robotic grasping as target-centric dependency reasoning, focusing on identifying and removing obstructions.
• It uses a directed obstruction graph and grounded chain-of-thought traces to determine the minimal set of objects to remove before grasping the target.
• The model combines RGB-D observations, reinforcement finetuning, and benchmark evaluations to enhance performance in both synthetic and real-world cluttered settings.

Searching arXiv for the primary UNOGrasp paper and closely related work on unified grasping, obstruction reasoning, and universal dexterous grasping. arXiv search query: "UNOGrasp obstruction reasoning robotic grasping" UNOGrasp is a vision-LLM for obstruction reasoning in robotic grasping, introduced together with UNOBench to study how a robot should infer the sequence of actions needed to unobstruct a target object before grasping it. Rather than treating grasping in clutter as a pure target-grounding problem, it formulates grasping as target-centric dependency reasoning over visible objects, where the robot must identify not only the referred object but also the currently accessible obstructors whose removal will expose it (Jiao et al., 28 Nov 2025).

1. Problem formulation and target-centric obstruction structure

UNOGrasp addresses cluttered robotic grasping under an RGB-D observation $I=(rgb,d)$, a visible object set $\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$, and a free-form language instruction $q \in L$ that refers to a target object $o_t$. The objective is not merely to localize $o_t$, but to infer an unobstruction plan: if the target is accessible, grasp it directly; otherwise, identify the minimal set of currently accessible obstructors whose removal will reduce obstruction and eventually expose the target (Jiao et al., 28 Nov 2025).

The central abstraction is a directed target-centric obstruction graph $G_t=(\mathcal{V}_t,E_t)$. An edge $(o_i,o_j)\in E_t$ means that $o_i$ is obstructed by $o_j$ from the camera viewpoint. Obstruction paths are directed chains that begin at the target and end at accessible top-level obstructors. The set of target ancestors is defined as

$$\mathcal{A}(o_t) {=} \{\, o_i {\in} \mathcal{O} \mid \exists \text{ a directed path } [o_t, {\cdots}, o_i] {\in} G_t \,\},$$

and the set of top-level accessible obstructors is

$\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$0

The model objective is written as

$\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$1

This formulation makes accessibility, rather than object identity alone, the primary prediction target (Jiao et al., 28 Nov 2025).

A notable feature of the method is that it treats obstruction reasoning as explicitly multi-step and target-centric. Prior graph-based approaches to clutter often reason over all pairwise relations in the scene, whereas UNOGrasp restricts reasoning to the dependency structure rooted at the referred object. This suggests a computational and representational bias toward action-relevant scene structure rather than exhaustive relational modeling.

2. Reasoning format, visual grounding, and model architecture

UNOGrasp is built on Qwen2.5-VL-3B-Instruct. It does not introduce a new low-level visual backbone; instead, it specializes a vision-LLM through a task-specific reasoning format and a training recipe centered on obstruction paths and grounded visual cues (Jiao et al., 28 Nov 2025).

The model is trained to produce reasoning traces in a structured output format with > ... followed by a final action set in <answer>...</answer>. In the natural-language setting, each reasoning step includes the object’s name or description, its image coordinates, the obstruction relation to the next object, and optionally the occlusion or obstruction ratio. The final action plan is the set $\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$2, namely the currently accessible top-level obstructors (Jiao et al., 28 Nov 2025).

A distinctive design choice is the inclusion of obstruction-aware visual cues in the reasoning traces. These cues include contact points, obstruction ratios, and degree words such as slightly, partially, mostly, and heavily obstructed. The paper reports that occlusion ratio is the most helpful auxiliary signal during supervised training. This implies that obstruction severity functions as a valuable intermediate representation even though the final decision target is a discrete action set (Jiao et al., 28 Nov 2025).

The perception-and-action interface is modular. RGB is used for reasoning and action-sequence planning, while depth is used downstream to estimate 3D grasp points. In the robotic system, GroundedSAM is used to obtain the target object mask, and GraspNet predicts candidate 6-DoF grasps; these execution backends are shared across compared reasoning agents so that measured differences are attributable to reasoning rather than grasp synthesis (Jiao et al., 28 Nov 2025).

The benchmark supports two settings. In the Oracle setting, objects are referred to only by Set-of-Mark IDs, which isolates reasoning because grounding ambiguity is removed. In the Natural Language Prompting setting, the target is referred to through free-form language, and the model must perform both grounding and obstruction reasoning. This separation is methodologically important because it distinguishes failures of language grounding from failures of dependency reasoning.

3. Training methodology and verifiable reasoning optimization

UNOGrasp is trained in two stages: supervised finetuning and reinforcement finetuning. The supervised stage uses examples $\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$3, where $\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$4 is the grounded reasoning trace and $\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$5 is the final answer. The corresponding loss is

$\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$6

This stage teaches the model to identify the referred target, distinguish unobstructed, single-path, and multi-path cases, traverse valid contacting obstruction chains, and output grounded names, coordinates, and top-level blockers (Jiao et al., 28 Nov 2025).

During supervised finetuning, vision tuning is frozen, while the MM-MLP and LLM components are tuned. Reported hyperparameters are: base model Qwen2.5-VL-3B-Instruct, 2 epochs, learning rate $\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$7, batch size per device 4, bf16 precision, cosine scheduler, warmup ratio 0.03, gradient clipping 1.0, max sequence length 8192, max image pixels 12.8M, and DeepSpeed ZeRO-3. Training used 4 A100-SXM-64GB GPUs (Jiao et al., 28 Nov 2025).

After supervised finetuning, the model is optimized with Group Relative Policy Optimization. The reward is a weighted combination of a format reward and a task reward: $\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$8 The format reward is binary and checks whether the output contains properly formed and closed <think> and <answer> sections. The task reward supervises the predicted top-obstructor set using set IoU: $\mathcal{O}=\{o_1,o_2,\ldots,o_N\}$9 Group-wise relative advantages are defined by

$q \in L$0

and the reinforcement loss is

$q \in L$1

Reported reinforcement hyperparameters are: 1 epoch, batch size per device 8, gradient accumulation 2, bf16 precision, group size $q \in L$2, max completion length 512, $q \in L$3, temperature 1.0, top-$q \in L$4, FlashAttention-2, gradient checkpointing enabled, and DeepSpeed ZeRO-2 (Jiao et al., 28 Nov 2025).

A key empirical claim is that rewards defined only on final answer sets improve not only answer accuracy but also internal reasoning quality. The paper reports improvement in a post hoc path-level fidelity metric as well. This suggests that verifiable set-level supervision is sufficient to shape intermediate chain-of-thought behavior in this domain, although the optimization target itself remains answer-based rather than path-based.

4. UNOBench: benchmark design, annotations, and evaluation metrics

UNOBench is introduced as a large-scale benchmark for linguistically grounded obstruction reasoning. It is built from MetaGraspNetV2 and augments it with free-form object descriptions and target-centric obstruction graphs with contact points, obstruction ratios, and obstruction degrees (Jiao et al., 28 Nov 2025).

The construction pipeline proceeds through four steps: Set-of-Marks preparation, obstruction information extraction, object-centric obstruction graph construction, and ID-name-coordinate association. GPT-4o is used to assign names and descriptions to object IDs, followed by human revision. The paper reports 5,400 challenging images reviewed, 41,193 object names checked, 196 native or primary English speakers recruited on Prolific, 80 minutes per annotator as target workload, 4,678 corrected images, and 17,261 revised object names. Annotators could also mark heavily occluded objects as “indescribable object” (Jiao et al., 28 Nov 2025).

The benchmark statistics reported in the main paper are shown below.

Split	Scenes / Images	Objects / Paths
Synthetic	6,255 scenes, 25,020 view images	97,066 named object instances, 108,174 reasoning paths
Real	520 scenes	2,232 named object instances, 2,552 obstruction paths

Synthetic scenes are split 7:1:2 into train, validation, and test, while real scenes are test only. Difficulty is defined by minimum graph depth $q \in L$5 and number of distinct reasoning paths $q \in L$6: No-Occ has $q \in L$7; Easy has $q \in L$8; Medium has either $q \in L$9 or $o_t$0; Hard has either $o_t$1 or $o_t$2 (Jiao et al., 28 Nov 2025).

UNOBench evaluates performance at three levels. Outcome-level metrics compare predicted and ground-truth top-obstructor sets using SR-P, SR-R, and SR-F1. Object-level reasoning metrics extract pairwise obstruction triplets and compute OP, OR, and $o_t$3. Path-level reasoning is measured by Multi-Path Normalized Edit Distance: $o_t$4 with Hungarian matching cost

$o_t$5

and

$o_t$6

Lower MP_NED indicates closer structural agreement between predicted and ground-truth obstruction paths (Jiao et al., 28 Nov 2025).

5. Empirical performance and robotic evaluation

On synthetic Oracle evaluation, UNOGrasp is reported as best across all difficulty levels. Path-level results include No-obstruction SR 94.8, Easy SR-F1 83.3, Medium SR-F1 69.1, Hard SR-F1 54.5, with MP_NED values 0.11, 0.37, and 0.51 for Easy, Medium, and Hard. In Natural Language Prompting, performance decreases as expected because grounding is harder, but UNOGrasp remains strongest, with No-obstruction SR 92.5, Easy SR-F1 74.9, Medium SR-F1 59.7, Hard SR-F1 37.2, and MP_NED 0.06, 0.20, 0.53, and 0.67 across No-Obs, Easy, Medium, and Hard (Jiao et al., 28 Nov 2025).

Object-level reasoning metrics show a similar pattern. In synthetic Oracle evaluation, UNOGrasp achieves Easy $o_t$7 82.6, Medium 62.0, Hard 48.7, and overall 75.3. In Natural Language Prompting, it achieves Easy 66.1, Medium 39.5, Hard 25.7, and overall 57.2. The paper notes that Gemini and Qwen baselines are mostly below 4 overall in this grounded natural-language setting, indicating that multi-step reasoning must remain tightly aligned to object identities and coordinates to be effective (Jiao et al., 28 Nov 2025).

On the real test set, trained only on synthetic data, UNOGrasp maintains strong performance. In path-level Oracle evaluation it reports No-obstruction SR 72.5, Easy SR-F1 77.2, Medium 64.4, Hard 63.9. In path-level Natural Language Prompting it reports No-obstruction SR 70.0, Easy 71.3, Medium 47.3, Hard 40.3. Object-level real Oracle overall $o_t$8 is 71.4, and natural-language overall $o_t$9 is 49.1 (Jiao et al., 28 Nov 2025).

The paper also reports ablations. Adding obstruction cues during supervised finetuning improves synthetic overall performance from SR-F1 74.7, OR-F1 71.9, MP_NED 0.220 to 76.4, 73.3, and 0.210 when using occlusion ratio. Reinforcement finetuning further improves Easy SR-F1 from 81.8 to 83.3, Medium from 67.1 to 69.1, Hard from 50.1 to 54.5, and overall from 76.4 to 78.2. Overall OR-F1 improves from 73.3 to 75.3, and MP_NED from 0.210 to 0.201 (Jiao et al., 28 Nov 2025).

Real-world robotic experiments use a UR5e arm with a Robotiq 2F-85 gripper and a ZED 2 top-down stereo camera at approximately 0.8 m, in a bin with 25 distinct household objects and 30 real-world scenarios spanning Easy, Medium, and Hard. A scenario succeeds only if the robot removes objects in the correct order and ultimately grasps the target. Reported success ratios are: Gemini Robotics-ER 1.5 at 80% / 30% / 10% for Easy / Medium / Hard and 40% average; Qwen2.5-VL at 10% / 0% / 0% and 3% average; UNOGrasp at 80% / 30% / 40% and 50% average. The most pronounced gain is on Hard scenarios, where UNOGrasp improves by 30 points over Gemini (Jiao et al., 28 Nov 2025).

6. Position within the grasping literature and limitations

UNOGrasp occupies a distinct position within the broader grasping literature. Earlier “unified” and “universal” systems such as UniGrasp (Shao et al., 2019), UniDexGrasp (Xu et al., 2023), UniDexGrasp++ (Wan et al., 2023), UniGraspTransformer (Wang et al., 2024), RobotFingerPrint (Khargonkar et al., 2024), and UniMorphGrasp (Wu et al., 31 Jan 2026) focus primarily on cross-object or cross-embodiment grasp synthesis and execution. Their central abstractions involve contact-point prediction, proposal generation, morphology conditioning, policy distillation, or shared gripper-coordinate spaces. UNOGrasp differs in that its primary bottleneck is not grasp synthesis itself, but target accessibility under clutter: it reasons about which objects must be removed before a grasp becomes possible (Jiao et al., 28 Nov 2025).

It is also distinct from open-vocabulary constrained grasping systems such as UniDiffGrasp, which integrate vision-language reasoning with part-conditioned diffusion for semantic part grasping (Guo et al., 11 May 2025). Both methods connect language and grasping, but their problem formulations differ: UniDiffGrasp constrains where to grasp on a semantically specified object part, whereas UNOGrasp reasons about which objects must be removed first so that the target becomes reachable. A plausible implication is that the two lines of work are complementary: obstruction reasoning determines the correct action sequence in clutter, while constrained grasp synthesis determines the correct contact geometry once the target or intermediate obstructor has been selected.

UNOGrasp is also related, at a higher level, to work on nonprehensile pregrasp for initially ungraspable objects, which synthesizes environmental-contact maneuvers to expose graspable surfaces (Chen et al., 2023). The difference is methodological. The nonprehensile pregrasp framework plans physically realistic object-hand trajectories through graph search, optimal control, and a learned graspability metric, while UNOGrasp reframes accessibility as a visually grounded reasoning problem over target-centric obstruction paths. This suggests two complementary notions of pre-grasp intelligence: symbolic obstruction planning and low-level nonprehensile trajectory synthesis.

The paper identifies several limitations. Obstruction severity is not modeled in execution planning; the plan is based only on whether an obstruction exists, not how difficult it is physically to remove. The reasoning is single-view and future work is suggested on multi-view perception. Depth is not integrated into the reasoning model itself; RGB drives reasoning, while depth is used later for grasp points. The output is a final action set rather than a complete manipulation policy, since grasp feasibility and motion planning are delegated to external modules. Reported failure cases include contact mistaken for obstruction, visually similar objects, dense clustered groups, hallucinated obstructions in no-obstruction scenes, and coordinate misalignment under natural-language prompting (Jiao et al., 28 Nov 2025).

Within this context, UNOGrasp’s main contribution is to establish obstruction reasoning as a first-class embodied reasoning problem for robotic grasping. Its central technical claim is that grasping in clutter requires reasoning over a target-centered dependency structure, and that such reasoning can be trained and evaluated at scale through explicit obstruction graphs, grounded chain-of-thought supervision, and verifiable reward signals (Jiao et al., 28 Nov 2025).