- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper uses a RL skill discovery-style motivation for VLAs. They use a VQ-VAE with some factorization to create independent skills. Then they train LORA adaptors on top of a VLA to output the actions of those skills. Lastly they train a router to choose the skill which yields the lowest prediction accuracy, with an additional regularizer for the VQ vae predicted skill. This yields a VLA with separate skills to be used in different environments.
Critical strengths
The idea of breaking VLA behaviors into more intrepretable skills is intuitive and I like the connection to RL. Using LORA to parameterize skills instead of the skill variable is an interesting idea. This allows them to preserve pretraining while also specializing towards skills.
Critical weaknesses
I dont find the experiments very convincing. They train on data from the same simulation as they use during evaluation. Even when the skills are from a different task set, they retrain the VLR. A better experiment would evaluate the model on a completely different sim, or sim2real, etc, to test if these learned skills generalize or not. Libero is pretty saturated so these results are not too interesting either.
Questions for authors
- How much does the second term in equation 11 matter? It seems unnecessary if the motivation in this section is correct.
- How does the adaptation cost compare to other methods? The training pipeline is extensive.
- Are you finetuning pi 0.5 on the new tasks since you finetune your model (the router)?
- How does your method compare to instead conditioning on the latent skill representation?
5 comments
Response to Reviewer 29BG (2/3). Continuation of our response to W1 (Part 2/2): held-out real-robot compositions and the observed limits of skill reuse and our response to Q1.
(c) Real-robot generalization. Both methods use only A/B demonstrations for SFT, discovery, skill fitting, and router training. C–J are held out from every stage; Table 3 identifies the reused grasp and placement components.
Table 3. Real-robot grasp-and-placement tasks and results. 10 rollouts per task per method; entries are success counts. The arrow denotes grasping followed by placement/stacking, e.g., “Pick up the yellow bag, then place it on the right.” Object goals specify stacking on top of that object. ✓/✗: seen/unseen in A/B; placement denotes the object–goal pair.
| ID | Task: grasp → place/stack | Grasp seen | Goal seen | Placement seen | π0.5 | Ours |
|---|---|---|---|---|---|---|
| A | Yellow bag → right | ✓ | ✓ | ✓ | 7 | 9 |
| B | Purple bag → on green box | ✓ | ✓ | ✓ | 8 | 10 |
| C | Yellow bag → on green box | ✓ | ✓ | ✗ | 3 | 7 |
| D | Green box → right | ✗ | ✓ | ✗ | 0 | 8 |
| E | Purple bag → right | ✓ | ✓ | ✗ | 4 | 8 |
| F | Yellow bag → on purple bag | ✓ | ✗ | ✗ | 0 | 5 |
| G | Purple bag → on yellow bag | ✓ | ✗ | ✗ | 0 | 4 |
| H | Yellow bag → left | ✓ | ✗ | ✗ | 0 | 0 |
| I | Purple bag → left | ✓ | ✗ | ✗ | 0 | 0 |
| J | Green box → left | ✗ | ✗ | ✗ | 0 | 0 |
PlugVLA raises success on held-out recombinations C/E from 30%/40% to 70%/80% and succeeds on D/F/G where π0.5 records zero successes. These tasks change an object’s role as the manipulated object or placement support, supporting motor-skill reuse under changed task contexts. Both methods fail on the unseen left-goal tasks H–J, delimiting the observed scope of reuse.
Q1. How much does the second term in equation 11 matter? It seems unnecessary if the motivation in this section is correct.
Response. The auxiliary term retains discovered motion structure alongside action-fit supervision. We test its contribution through success and boundary-action ablations, then relate it to motion retention during deployment.
(a) Complementary supervision targets. groups demonstrations for skill fitting; selects the fitted update with the lowest action loss in context. The auxiliary term connects context-dependent selection to the discovery grouping; Table 4 tests its practical contribution.
(b) Supervision ablation. Table 4 compares action-fit-only, discovery-label-only, and mixed supervision; mixed uses auxiliary weight 0.1. The base policy and skill bank remain frozen during router training.
Table 4. Router-supervision ablation: SR (%) and boundary action change. 100 rollouts per task; boundary statistics cover all 8 tasks (800 per method). Boundary change is the mean normalized RMS change in arm commands across chunks on RoboTwin8 Hard, excluding grippers; lower indicates smoother transitions.
| Supervision | Objective | SR (%): RoboTwin8 Easy | Hard | Mean boundary action change |
|---|---|---|---|
| FM-only | 59.2 | 29.3 | 0.033 | |
| VQ-only | 58.6 | 28.7 | 0.025 | |
| Mixed | 61.9 | 32.9 | 0.027 |
Mixed supervision achieves the highest SR, gaining 2.7/3.6 pp over FM-only on RoboTwin8 Easy/Hard and reducing boundary change from 0.033 to 0.027. VQ-only yields the smallest change (0.025) but lower SR. Combining the targets therefore improves success while retaining smoother transitions than action-fit supervision alone.
(c) Motion retention during execution. On RoboTwin8 Hard, predicted skills match executed-motion labels in 89.34% of successful rollouts versus 46.78% of failed ones (see our response to Reviewer TYMh, Q2). This associates successful execution with retained motion structure. The ablation tests the auxiliary term’s effect; the agreement analysis supports the relevance of its motion-based target.
The auxiliary term guides skill prediction, rather than directly penalizing adjacent action differences; boundary changes are measured consequences of the resulting policy.
Response to Reviewer 29BG (3/3). Q2,3,4.
Q2. How does the adaptation cost compare to other methods? The training pipeline is extensive.
Response. We report preparation time and added parameters, compare the stages with related training pipelines, and explain bank reuse. Candidate scoring is performed offline.
(a) Measured preparation time. On one 80 GB A100, RoboTwin8 skill fitting, scoring, and router training each use 3 epochs and take 4.9/2.6/3.8 h (11.3 h total); base SFT takes 54 h over 6 epochs. These stages fit local specializations, construct action-fit labels, and learn skill prediction.
Candidate scoring is offline. Deployment predicts a skill and executes its update through the shared policy, without candidate-wise loss evaluation.
(b) Parameter overhead. Table 5 includes frozen additions and excludes base weights and routing/retrieval heads, separating parameter storage from the measured preparation time.
Table 5. Additional model parameters.
| Method | Added parameters (M) |
|---|---|
| KinRT-LoRA | 1,076.724 |
| BehaviorVLA | 33.627 |
| PlugVLA | 47.481 |
(c) Related training pipelines. Related methods also construct supervision and organize training into stages:
- KinRT clusters action trajectories and uses cluster IDs to supervise an observation-based expert router.
- ChatVLA-2 [1] augments reasoning annotations with GPT-4o, co-trains on image-text and robot data, then freezes the VLM to train the action expert.
- InstructVLA [2] pretrains an action expert, then freezes it for VLA instruction tuning with MoE adaptation, using GPT-4o-assisted annotations.
PlugVLA similarly constructs supervision before learning its predictor, using existing demonstrations to fit parameter skills, score their action fit, and train the VLR. These stages separate skill content from skill selection, enabling reuse of the fitted bank.
(d) Preparation and reuse. The additional preparation builds a reusable skill bank. In cross-task reuse, its parameters remain fixed while the base receives target-task SFT and the router adapts selection.
Q3. Are you finetuning pi 0.5 on the new tasks since you finetune your model (the router)?
Response. Yes, π0.5 receives target-task SFT, and PlugVLA uses the same reference policy.
Q4. How does your method compare to instead conditioning on the latent skill representation?
Response. We compare latent skill embeddings with parameterized skills at approximately matched total adaptation capacity, directly testing the representation choice.
Latent baseline implementation. The baseline uses 16 learned skill embeddings and layer-specific FiLM modulation before the action-expert FFNs, with the SFT backbone frozen. It shares our segment assignments, flow-matching objective, and router architecture, with its own recomputed routing labels; total adaptation capacity and fitting exposure are matched, but active capacity and FLOPs differ.
(a) Matched-capacity results. PlugVLA uses 47.481M adaptation parameters versus 47.493M for latent embeddings, excluding the router. It improves LIBERO from 97.1% to 98.0% (+0.9 pp) and RoboTwin8 Hard from 27.0% to 32.9% (+5.9 pp). Evaluation uses 50 rollouts per LIBERO task and 100 per RoboTwin8 task.
(b) Representation choice. The gains support encoding recurring behavior as detachable parameter updates. Each update specializes gated feature formation and residual modulation while preserving the shared policy’s native action interface; the router predicts its activation from task context.
References
[1] Z. Zhou et al. ChatVLA-2: Vision-Language-Action Model with Open-World Embodied Reasoning from Pretrained Knowledge. NeurIPS, 2025.
[2] S. Yang et al. Vision-Language-Action Instruction Tuning: From Understanding to Manipulation. ICLR, 2026.
Response to Reviewer 29BG (1/3). Response to W1 (Part 1/2): broader-task evaluation, frozen-module transfer, and VLR-only adaptation.
Response to Reviewer 29BG
Thank you for the feedback and for recognizing the connection to skill learning and the use of low-rank parameter updates to preserve the shared policy while specializing local motor behavior.
Response. Our contribution is to represent recurring local motions as independently activatable parameter skills, separating skill content from selection. Tables 1–3 test performance on harder tasks, frozen-module reuse, robustness under LIBERO-Plus perturbations, and held-out real-robot compositions.
(a) Broader task coverage. Table 1 compares the methods on the remaining 33 RoboTwin2.0 Hard tasks, with three representative examples.
Table 1. SR (%) on 33 additional RoboTwin2.0 Hard tasks. 100 rollouts per task.
PlugVLA improves average SR from 41.45% to 48.08% (+6.63 pp). Beat Block Hammer and Open Microwave gain 10 and 18 pp. This extends evaluation beyond the original task groups and the high-success LIBERO setting.
(b) Transfer and VLR adaptation. RoboTwin transfers the bank and router between the 8/9-task groups without updating either module; the base still receives target-task SFT. LIBERO-Plus evaluates robustness on LIBERO tasks across seven perturbation categories: object layout, camera viewpoint, robot initial state, language, lighting, background, and sensor noise. We distinguish direct evaluation of LIBERO-trained models from additional VLR-only training on LIBERO-Plus demonstrations.
Table 2. Transfer and VLR-only adaptation: SR (%).
Table 2a. RoboTwin transfer. PlugVLA's source skill bank and VLR remain frozen; both methods use target-task SFT. 100 rollouts per task.
Table 2b. LIBERO-Plus evaluation. All models start from training on original LIBERO data. π0.5 and frozen-VLR PlugVLA use no LIBERO-Plus training data. Adapted-VLR PlugVLA additionally trains only the VLR on LIBERO-Plus demonstrations, keeping the base policy and skill bank frozen.
With both modules frozen, PlugVLA improves RoboTwin transfer by 2.1/1.4 pp and LIBERO-Plus by 0.5 pp on average. On LIBERO-Plus, five of seven categories improve, while language and sensor noise decline by 1.0/1.8 pp.
VLR-only adaptation raises the LIBERO-Plus average to 87.3%, gaining 2.2 pp over the frozen VLR and 2.7 pp over π0.5. All seven categories improve over the frozen VLR, led by camera/sensor (+4.6/+4.5 pp). This supports adapting skill selection to the target distribution while reusing unchanged skill parameters.