- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
TriRelVLA replaces dense visual tokens in a VLA with an explicit object-hand-task relational representation: object tokens are grounded from fused SigLIP/VGGT features, a hand token is anchored on proprioception, and the instruction is decomposed into action, role, constraint and stage tokens. A task-guided graph with typed edges (task-object, task-hand, object-hand, object-object) is updated by a relation-aware graph transformer, and a bottleneck projects it into a Qwen3-4B LLM for action prediction. Training is three-stage (OXE, DROID, LIBERO + CSOT-Bench) with object/hand mask auxiliary losses. The authors also introduce CSOT-Bench, a real-robot benchmark with cross-scene, cross-object and cross-task suites. Against eight VLAs ranging from OpenVLA to SemanticVLA, the method is at parity on fine-tuned LIBERO/CSOT (97.6/84.9) and reports large gains on the CSOT generalization suites (e.g., cross-object 80 vs. 47).
Critical strengths
- The triadic object-hand-task representation is a well-motivated architectural idea. It moves beyond scene-level structured representations to action-centric relations, and the implementation - task-guided nodes, a relation-aware graph transformer, and a relational bottleneck - is coherent and clearly specified. The ablation trends (triadic representation, graph, proprioception, task grounding) all point in the same direction, although, as noted below, their magnitudes are not yet reconcilable. Fine-tuned performance is competitive with strong VLA baselines (97.6 LIBERO, 84.9 CSOT). This is consequential because it offers a concrete alternative to dense visual tokens for generalization.
- The paper reports large cross-scene/object/task gains on a real-world benchmark. On CSOT-Bench, TriRelVLA improves over SemanticVLA by 21, 33, and 25 points on the first task of each suite (78 vs. 57 cross-scene, 80 vs. 47 cross-object, 71 vs. 46 cross-task), while remaining comparable on fine-tuned tasks. If these gains hold under a controlled evaluation, they would be significant for real-world VLA deployment.
Critical weaknesses
- The ablation is not interpretable as presented. The "Generalized task" column in Tables 2-6 is never defined, and 79.5 for the full model cannot be reconciled with Fig. 7 (79/81/73 cross-scene/object/task), so it is unclear what the column aggregates. The jump from 43.0 to 79.5 when adding the relational graph (+36.5) is unexplained, while the gain over the best baseline is only +0.7; no per-suite results, parameter counts, or training controls are given, so the graph may be confounded with capacity or training. The grid is also incomplete: Table 3 has no pairwise combinations (object+hand, object+task, hand+task), Table 4 does not isolate L_obj and L_hand, Table 6 omits the cell with object grounding off and hand grounding on, and no table reports error bars, seeds, or the number of rollouts. These gaps are consequential because the ablation is the main evidence for the architectural claims.
- The comparison is selective and not controlled for model size, training data, or compute. Table 1 compares eight baselines, but Figures 6-7 compare only four (OpenVLA, Octo, CogACT, SemanticVLA), omitting SpatialVLA (CSOT 80.8) and CogVLA (82.5), the strongest structured baselines in Table 1, while keeping Octo (72.0). The paper reports no parameter counts, pretraining data, or compute for the baselines, though TriRelVLA uses a 4B Qwen3 LLM with SigLIP/VGGT and a three-stage OXE/DROID pipeline, whereas Octo is ~0.1B and DiffusionPolicy is trained from scratch. The large gains are on CSOT-Bench, a self-constructed benchmark that is also used for Stage-3 training and that is described without task, demonstration, object, or split counts, and for which only code release is promised. This is consequential because the paper does not isolate the claim that the triadic structure, rather than extra data or capacity, drives generalization.
Questions for authors
- The relational graph adds +36.5 generalization (43.0→79.5) but only +0.7 fine-tuned (84.2→84.9). What explains this asymmetry? Please report a controlled ablation with the same data, training stages, parameter count, and compute. The basic architecture’s generalized score (32.1) is close to OpenVLA’s generalized average (~34.5 in Figure 7), so the surprising result is the full model’s 79.5, which is far above every baseline (SemanticVLA ~49.7). What causes this jump?
- The comparison lacks several recent related works, such as GraphCoT-VLA (Huang et al., 2026) and OpenVLA-OFT (Kim et al., 2025). Can you compare against them, or explain why they are not applicable?