PaperlayerView exact revision on arXiv

Source-linked AI summary

Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment

Dwip Dalal, Shivansh Patel, Chahit Jain, Jeonghwan Kim, Utkarsh Mishra, Alex Baratian, Hyeonjeong Ha, Heng Ji, Svetlana Lazebnik, Unnat Jain

arXiv:2607.13429v1cs.ROcs.CV

TL;DR

Behavior-cloning finetuning can overwrite pretrained VLM representations and leave language and action predictions misaligned, limiting VLA generalization. Anchor-Align preserves those representations and jointly trains language and action on robot observations, nearly doubling real-robot task success while improving simulation generalization.

  • Problem

    Behavior-cloning finetuning overwrites pretrained visuolinguistic representations, while co-training leaves language and action heads supervised on separate observations.

  • Method

    Anchor-Align combines representation distillation from a frozen VLM copy with motion-direction language supervision on the same observations used for action prediction.

  • Results

    Anchor-Align consistently improves OOD, perceptual-robustness, and long-horizon performance in simulation and raises mean real-robot success from 28.3% to 54.2% and 36.7% to 60.0%.

  • Takeaways & Limitations

    Effective VLA finetuning can preserve pretrained semantics while grounding them in action, using supervision already contained in robot demonstrations.

Abstract

from arXiv · show

Finetuning a pretrained vision-language model (VLM) on robot demonstrations via behavior cloning (BC) has become the standard recipe for vision-language-action (VLA) policies. However, BC finetuning progressively overwrites the pretrained representations that support visual and semantic generalization. Co-training on web image-text data, a common remedy, does not prevent this; it applies language and action losses to separate observations, leaving VLAs with language-action misalignment that standard manipulation benchmarks do not expose. We propose Anchor-Align, which augments BC with two objectives: Vision-Language Anchoring distills layer-wise representations from a frozen VLM copy to prevent this drift, while Language-Action Alignment converts each action target into a discrete motion-direction label and jointly trains language and action prediction on the same robot observation. On a physical xArm7 robot, across two widely used VLA architectures, Anchor-Align improves real-robot success on both (28% to 54% and 37% to 60%). At scale in simulation, we demonstrate consistent improvements on OOD perturbations, perceptual robustness, and long-horizon control across LIBERO-PRO, LIBERO-Plus, and CALVIN, respectively, suggesting that preserving pretrained representations and effective action learning are not fundamentally at odds. Project page: anchoralignvla.github.io

1 Introduction

The introduction identifies two failures of behavior-cloning finetuning: it erases pretrained vision-language concepts and leaves language and action predictions misaligned. Anchor-Align addresses these problems by anchoring layer-wise VLM representations and jointly training language-action alignment on robot observations alongside BC.

  • Failure mode: representation drift: Standard BC optimizes only action prediction, allowing finetuning updates to overwrite pretrained visuolinguistic and spatial representations.The introduction attributes this drift to the absence of an objective protecting the VLM’s pretrained representations.
  • Proposed method: Anchor-Align combines Vision-Language Anchoring, Language-Action Alignment, and standard BC to preserve VLM priors while aligning language with action.Anchoring distills representations from a frozen pretrained VLM at every transformer layer, while alignment trains a discrete motion-direction label on the same observation used for action prediction.
  • Motivation: 90% of the time, the xArm7 VLA reaches for the green mug when instructed to pick up the pink one, indicating erased color grounding.The example illustrates how BC finetuning can overwrite pretrained visuolinguistic representations needed for out-of-distribution generalization.
  • Failure mode: language-action misalignment: Co-training supplies disjoint supervision to language and action heads, so predictions on the same robot observation may contradict each other.For example, the action head may predict movement to the right while the language head predicts “left.”
  • Evaluation: Anchor-Align improves OOD generalization on LIBERO-PRO and LIBERO-Plus under unseen spatial rearrangements, camera perturbations, and other training-absent shifts.The method is evaluated under both regression and flow-matching action heads, as well as in simulation and real-world experiments.

2 Related Works

Prior work develops general-purpose VLAs by adapting pretrained vision-language models to robot manipulation, while addressing representation forgetting through robot and generic vision-language co-training. However, existing evaluations do not test whether language and action predictions remain consistent on the same observation.

  • General-Purpose VLAs for Robot Manipulation: General-purpose VLAs adapt pretrained vision-language models to downstream manipulation tasks with minimal finetuning.RT-2 co-finetunes on robot trajectories, and OpenVLA extends this through cross-embodiment training on Open X-Embodiment.
  • Preserving VLM Representations in VLAs: Pretrained representations support manipulation performance, motivating co-training VLAs on robot and generic vision-language data to mitigate catastrophic forgetting.This preservation requirement applies to the full VLM backbone because semantic and spatial reasoning span vision and language representations.
  • Embodied Question Answering: Co-trained VLAs evaluate language understanding on generic benchmarks but do not test consistency between language and action predictions for the same observation.This gap distinguishes embodied question answering from the language-action consistency required in VLA evaluation.

3 Anchor-Align Method

Anchor-Align finetunes a pretrained VLM into a continuous-action VLA while preserving pretrained vision-language representations. It adds layer-wise representation anchoring and same-observation language-action alignment to standard behavior cloning.

  • Base VLA and objectives: The VLA combines a pretrained vision-language backbone and dedicated action head with proprioceptive conditioning through a lightweight state projector.Learned action queries extract task-relevant information from multimodal representations for action prediction.
  • Base VLA and objectives: Training uses standard action behavior cloning plus anchoring and alignment losses, with the BC form determined by the action-head architecture.The action loss is L1 for regression heads and L2 over predicted velocity for flow-matching heads; both added objectives use the same robot observation.
  • Vision-Language Anchoring: A frozen copy of the pretrained VLM distills layer-wise representations into the trainable backbone to prevent catastrophic forgetting.The anchor VLM processes the same image-text batches in parallel and acts as a frozen teacher.
  • Language-Action Alignment: Continuous action targets are converted into discrete motion-direction labels, supervising the VLM backbone on the same observation used for continuous action prediction.Non-stationary chunks receive one of six labels—up, down, left, right, forward, or backward—based on averaged translation, dominant axis, and sign.

4 Experiments

Anchor-Align is evaluated across simulation and real-robot benchmarks, where it improves visual grounding, perturbation robustness, long-horizon control, and real-world manipulation. Ablations and representation analyses show that both anchoring and language-action alignment contribute, while anchoring preserves pretrained VLM reasoning.

  • Simulation benchmarks: Anchor-Align VLA outperforms all baselines on every LIBERO-PRO and LIBERO-Plus axis.The reported 5-seed means and variances confirm that these gains are statistically significant.
  • Simulation benchmarks: 22.6% on LIBERO-PRO position-swap exceeds VLA-Adapter’s 2.3% and the 0% scores of MolmoAct and OpenVLA-OFT.Position-swap rearranges task-relevant objects, testing whether policies ground actions in the observed scene rather than memorized scene-to-action mappings.
  • Simulation benchmarks: +8.9 background texture, +7.4 sensor noise, +6.5 robot initial state, and +5.8 lighting are Anchor-Align’s largest gains over VLA-Adapter on LIBERO-Plus.LIBERO-Plus evaluates robustness to changed textures, sensor noise, unseen initial arm configurations, and other perturbations.
  • Ablations: Each anchoring and alignment term independently improves standard BC on every reported axis, while combining them in Anchor-Align VLA gives the best performance.The ablation compares VLA-Adapter, Anchor VLA with layer-wise MSE distillation, Align VLA with language-action alignment, and the full method.
  • Real-world experiments: 54.2% mean success with VLA-Adapter and 60.0% with StarVLA demonstrate improved real-world generalization across architectures.With StarVLA, compositional object-layout rises from 35% to 50%, spatial rearrangement from 30% to 60%, and cluttered scenes from 45% to 70%.
  • Representation analysis: 71.9% GQA accuracy for Anchor-Align exceeds 61.0% for standard BC and 43.1% for the frozen VLM, indicating preserved vision-language reasoning alongside action performance.GQA measures compositional object, attribute, and spatial understanding during finetuning.

5 Conclusion and Future Work … B.1 Language-Action Alignment Head

Anchor-Align preserves pretrained VLM semantics while jointly grounding language and action, nearly doubling real-robot success and improving position-swap generalization. The appendix details implementation, baselines, extended evaluations, representation analyses, diagnostics, and the language-action alignment head, while suggesting extensions to video world models.

  • 5 Conclusion and Future Work: Anchor-Align preserves pretrained representations and jointly supervises language and action with motion-direction labels, nearly doubling real-robot task success.The approach addresses semantic overwriting and language-action misalignment using supervision already contained in robot demonstrations.
  • 5 Conclusion and Future Work: Anchoring and alignment may extend to action-conditioned video world models through frozen-teacher regularization and action-outcome consistency objectives.The proposed analogy includes recovering the conditioning action from a generated transition with an inverse-dynamics model.
  • Appendix: The appendix organizes reproducibility materials, including baseline details, implementation settings, training costs, extended quantitative results, representation analyses, diagnostics, and real-world and simulated rollouts.It also lists qualitative LIBERO results and multi-seed, robustness, CALVIN, CKA, action-decoding, and language-preservation analyses.
  • A Baseline Details: The appendix compares non-VLM, standard VLM-based, co-trained, and knowledge-preservation VLA baselines across standard LIBERO and related evaluations.Listed methods include Diffusion Policy, π0-FAST, SmolVLA, VLA-0, MolmoAct, ChatVLA, Magma, π0.5-KI, and VLA-Adapter[Frozen].
  • A Baseline Details: Long-horizon CALVIN comparisons additionally include ReconVLA, RoboFlamingo, and DeeR-VLA alongside the baselines introduced in the main paper.These methods target long-horizon, instruction-conditioned manipulation under environment transitions.
  • B Implementation Details: The implementation appendix provides full architectural specifications, hyperparameters, and training-cost breakdowns to improve reproducibility.This section frames the subsequent details on the alignment head and training setup.
  • B.1 Language-Action Alignment Head: The alignment head adds a single d = 896 projection with 803,712 parameters and uses the pre-action hidden state with a frozen language head to predict motion directions.The six labels are forward, backward, left, right, up, and down; near-stationary samples with τ = 0.15 are masked, and gradients flow through the projection and LoRA parameters.

B.2 Training Setup

The training setup uses four NVIDIA GH200 GPUs with bfloat16 mixed precision and PyTorch 2.7. Models receive two camera views plus proprioceptive state and use LoRA-based parameter-efficient finetuning with specified rank and scaling.

  • Training uses 4 NVIDIA GH200 GPUs with bfloat16 mixed precision and PyTorch 2.7.
  • The backbone VLM takes third-person and wrist-camera images together with the proprioceptive state as input.
  • LoRA finetuning uses rank r=64 and scaling factor α=128 across all backbone linear layers, with periodic merging into the base model.Training runs for 10,000 gradient steps.

B.3 Training Cost Comparison … C.2 Per-Suite Robustness Breakdowns

The extended results compare Anchor-Align’s training overhead and robustness benefits, while testing whether language-action alignment contributes beyond regularization and whether gains transfer across standard LIBERO suites.

  • B.3 Training Cost Comparison: Anchor-Align preserves pretrained visuolinguistic priors with 3.4× less overhead per step, no external data, and 7× less additional GPU memory than co-training.Table 6 frames the comparison against Standard BC and Co-training + KI as a cost-performance trade-off.
  • B.3 Training Cost Comparison: Co-training + KI incurs +95% wall-clock overhead, +5 GB GPU memory, and requires 25K external VQA samples, yet reaches 54.0% on LIBERO-PRO language rephrase.It also requires maintaining the language modeling head alongside the action head during training.
  • B.3 Training Cost Comparison: Anchor-Align uses one inference-only forward pass through a frozen VLM copy, replacing its language head with nn.Identity() and reducing the anchor footprint to 0.7 GB.The visible passage also reports +28% wall-clock overhead, or 0.36 s per step, and a 2,500-step run completed in 68 minutes.
  • B.4 Real-World Implementation: The real-world evaluation uses a wrist-mounted and external front-view RGB camera on a UFactory xArm7, with 150 teleoperated demonstrations collected at 30 Hz.The tasks require object discrimination, spatial reasoning, and instruction-conditioned target selection under clutter.
  • B.5 Alignment vs. Regularization: Shuffle and Scatter Controls: Shuffle and Scatter preserve Anchor-Align’s loss structure while removing genuine motion semantics by changing only the motion-class-to-target-token lookup table.Shuffle uses a fixed derangement of six direction words, whereas Scatter maps the six classes to meaningless words selected as mutually distant in frozen output-embedding space.
  • B.5 Alignment vs. Regularization: Shuffle and Scatter Controls: 22.6% LIBERO-PRO position-swap success is achieved by Anchor-Align VLA, while Shuffle and Scatter recover only the baseline’s near-zero performance.The position-swap axis tests re-grounding language to a perturbed scene rather than memorizing scene-to-action mappings.
  • C.1 Success Rates on Standard LIBERO Suites: Anchor-Align VLA achieves the highest success rate on most of the four standard LIBERO suites—Spatial, Object, Goal, and Long—surpassing prior methods including π0.5-KI and OpenVLA-OFT.These methods use substantially larger backbones and large-scale robotic pretraining, and the result extends gains beyond the OOD stress tests to in-distribution benchmarks.
  • C.2 Per-Suite Robustness Breakdowns: Across LIBERO Long, Object, and Goal, the extended breakdowns test whether the gains from LIBERO-Spatial carry over to long-horizon composition, object knowledge, and goal-related robustness.The Long-suite figure reports largest gains of +20.8% on Lighting Condition, +18.6% on Object Layout, and +17.7% on Camera Viewpoint.

C.3 Multi-Seed Evaluation: Statistical Significance · C.4 Full CALVIN Comparison

Anchor-Align’s improvements remain statistically significant across five seeds and transfer to a larger StarVLA backbone. In full CALVIN ABC→D evaluation, it outperforms every listed baseline at every chain length and achieves the longest average rollout.

  • C.3 Multi-Seed Evaluation: Statistical Significance: Seed-to-seed variability is at most 0.7 percentage points and substantially smaller than Anchor-Align VLA’s improvements over VLA-Adapter across LIBERO evaluations.The comparison uses five independent training seeds and reports mean ± standard deviation.
  • C.3 Multi-Seed Evaluation: Statistical Significance: On the LIBERO Object Suite, Anchor-Align VLA consistently covers a larger evaluation area, with the largest gain on Robot Init State at +18.6%.The figure excludes the Standard and Position Swap axes.
  • C.3 Multi-Seed Evaluation: Statistical Significance: On the LIBERO Goal Suite, Anchor-Align VLA’s largest gains are +11.9% on Language Instruction and +4.3% on Background Texture.The figure excludes the Standard and Position Swap axes.
  • C.3 Multi-Seed Evaluation: Statistical Significance: Anchor-Align VLA improves StarVLA from 73.6% to 89.0% on language rephrase and from 89.8% to 91.2% on object swap.These gains use a larger Qwen2.5 3B VLM with an OFT MLP-based action head.
  • C.3 Multi-Seed Evaluation: Statistical Significance: The larger-backbone results indicate that anchoring and alignment benefits are properties of the training objective rather than a particular VLM scale or action-head design.This conclusion follows the qualitatively consistent LIBERO-PRO improvements observed with the larger StarVLA configuration.
  • C.4 Full CALVIN Comparison: In the full CALVIN ABC→D comparison, Anchor-Align VLA outperforms all listed CALVIN baselines at every chain length.Table 9 reports the complete baseline set, with Len denoting average rollout length.
  • C.4 Full CALVIN Comparison: Anchor-Align VLA also attains the longest average rollout in the full CALVIN ABC→D comparison.The table identifies the best result in each column and distinguishes the second-best result.

D Representation Analysis: Language Preservation and Action Decodability … D.5 Direction Understanding in the Backbone

The representation analysis evaluates language preservation with text-token CKA and action information with linear-probe R2, then examines layer-wise language retention and motion-direction decodability. Anchor-Align is presented as improving action and directional information while preserving pretrained representations.

  • D Representation Analysis: Language Preservation and Action Decodability: The appendix introduces the evaluation metrics and representation-level analyses used to assess language preservation and action decodability.These analyses support the paper’s study of how finetuning affects pretrained internal representations.
  • D.1 Centered Kernel Alignment (CKA): Text-token CKA measures the geometric similarity between pretrained and finetuned VLM representations, directly reflecting preserved language understanding.At each transformer layer, CKA compares frozen pretrained text-token hidden states with finetuned backbone states across evaluation samples.
  • D.2 Action Decodability (Linear Probing R2): Linear-probe R2 measures how much ground-truth discretized action information is linearly accessible from mean-pooled hidden states at each layer.The actions are represented in R7 with 256 bins per dimension, and higher R2 indicates greater linear decodability without the action head.
  • D.3 Language Preservation and Action Decodability: The analysis compares training paradigms along language preservation and action decodability, reporting that standard BC catastrophically destroys pretrained text representations.Language preservation uses text-token CKA, while action decodability uses linear probing of backbone hidden states for ground-truth actions.
  • D.4 Layer-wise Language Preservation: Layer-wise CKA shows that standard BC progressively destroys pretrained text representations with depth, collapsing in the output layers.The measurement asks whether the observed GQA collapse reflects behavioral change alone or reshaping of the backbone’s internal representations.
  • D.4 Layer-wise Language Preservation: Anchor-Align VLA attains the highest action decodability while sustaining high language preservation, gaining action information without overwriting pretrained geometry.This relationship is summarized in Figure 13 as action decodability versus language preservation.
  • D.5 Direction Understanding in the Backbone: Layer-wise distillation preserves pretrained text representations across all layers, whereas standard BC collapses them in the output layers.This comparison is shown by per-layer text-token CKA between finetuned backbones and the pretrained VLM.
  • D.5 Direction Understanding in the Backbone: Anchor-Align VLA substantially improves motion-direction decodability on spatial and object splits, with the largest relative improvement on the OOD object split.The analysis linearly probes motion-direction concepts from backbone hidden states, while Table 10 reports linear-probe accuracy.

E Language-Action Diagnostic: Dataset Construction … E.4 Grasp Axis

The diagnostic uses a programmatic framework to extract language labels from robot trajectories across motion direction, task completion, orientation, and grasp axes. It applies axis-specific labeling, filtering, and sampling procedures to construct evaluation data from diverse manipulation episodes.

  • E Language-Action Diagnostic: Dataset Construction: The framework extends beyond motion direction to orientation change, grasp state, and task completion status, while the main paper focuses on motion direction.It is applied to the MolmoAct mid-training dataset, spanning 73 manipulation tasks with 7-DoF action vectors.
  • E.1 Direction Axis: 10,000 direction frames are assigned six labels—left, right, up, down, forward, and backward—with downward approaches most common at 32.1% and backward motion rarest at 5.1%.Direction labels are generated by mapping chunked actions to motion words through average chunking, filtering, and discretization.
  • E.3 Orientation Axis: 1,001 orientation frames are uniformly distributed across seven classes: no rotation, roll-cw/ccw, pitch-up/down, and yaw-left/right.Orientation uses a W = 7-frame sliding window, discards changes below µnoise ≈0.57°, and labels changes below µrot ≈1° as no rotation.
  • E.2 Task Completion Axis: Task completion is labeled from the last gripper-state transition: frames before it are incomplete, while frames at or after it are complete.The gripper command uses θg = 0.5, with values below the threshold indicating open and higher values indicating closed.
  • E.4 Grasp Axis: Grasp episodes are retained only when they contain both a first open-to-closed pick event and a subsequent closed-to-open place event.Qualifying episodes are partitioned into pre-grasp, grasp-to-release, and post-release stages.
  • E.4 Grasp Axis: Ns = 10 frames are sampled from each grasp stage, yielding up to 3Ns = 30 labeled frames per episode and an approximately 2:1 open-to-closed distribution.The imbalance arises because pre-grasp and post-release stages predominantly contain open-gripper frames.

F Quantification of Misalignment in SOTA Models · G Real-World Rollouts

The diagnostic evaluates language-action misalignment in three co-trained VLAs on real robot images, finding that language and action heads often achieve moderate or strong individual accuracy while agreeing substantially less across four axes.

  • F Quantification of Misalignment in SOTA Models: ChatVLA, MolmoAct, and Magma are evaluated per frame on real robot images using Language, Action, and Alignment accuracies.The evaluation uses the MolmoAct mid-training dataset because it contains real robot images and avoids placing the pretrained models out of distribution.
  • F Quantification of Misalignment in SOTA Models: Prior co-trained VLAs are poorly aligned across Direction, Task Completion, Grasp, and Orientation despite their separate head accuracies.Alignment is the fraction of frames on which language and action heads agree.
  • F Quantification of Misalignment in SOTA Models: Across Task Completion, Grasp, and Orientation, joint alignment is substantially lower than the moderate accuracy achieved by individual language and action heads.MolmoAct shows the largest per-head-versus-joint discrepancy across these additional axes.
  • F Quantification of Misalignment in SOTA Models: 7.4–20.9% alignment scores occur on Direction and Orientation, the lowest-scoring axes across all models.These fine-grained spatial axes reflect largely decoupled translational and rotational planning pathways.
  • F Quantification of Misalignment in SOTA Models: 41.3% action accuracy versus 22.6% language accuracy on MolmoAct’s Direction axis yields only 15.5% alignment.The action result exceeds the 16.7% chance level for 6-way classification, yet the two heads agree on only 15.5% of timesteps.
  • F Quantification of Misalignment in SOTA Models: 65.6% language accuracy and 44.3% action accuracy produce only 48.1% alignment for ChatVLA on Task Completion.The gap indicates that correct predictions from one head frequently co-occur with incorrect predictions from the other.
  • F Quantification of Misalignment in SOTA Models: 77.1% action accuracy and 51.8% language accuracy on MolmoAct’s Grasp axis yield only 37.8% alignment.The action head has learned reliable grasp control that the language channel cannot consistently describe.

G.1 Faster Task Completion in Real World · G.2 Object-Orientation Perturbation · G.3 Anchor-Align Real-World Rollouts

Anchor-Align VLA completes successful real-world pick-and-place rollouts 1.7× faster than standard BC, with more decisive actions and tighter completion-time variation. It also handles object-orientation, object-position, distractor, and clutter perturbations across real-world rollouts while preserving the task instruction and target identity.

  • G.1 Faster Task Completion in Real World: 1.7× faster average successful rollout completion distinguishes Anchor-Align VLA from standard BC on the xArm7 pick-and-place task.Anchor-Align also exhibits a tighter distribution and lower variance in completion times.
  • G.1 Faster Task Completion in Real World: Standard BC produces smaller, tentative action vectors near grasp points, often undershooting displacement and requiring repeated corrective adjustments.These low-magnitude predictions delay secure grasp acquisition.
  • G.1 Faster Task Completion in Real World: Anchor-Align VLA produces higher-magnitude, decisive grasp actions and continues transport and placement through direct trajectories.The resulting behavior is more efficient and natural-looking, with practical implications for real-world cycle time and throughput.
  • G.2 Object-Orientation Perturbation: Six broccoli orientations vary the target’s silhouette and visible structure while keeping its identity and language instruction fixed.This Compositional Object-Layout regime prevents matching a memorized canonical view and forces perception to be re-grounded in each rollout.
  • G.3 Anchor-Align Real-World Rollouts: All six broccoli rollouts succeed when object position and surrounding distractors are simultaneously swapped across scene configurations.These results indicate robust handling of joint object-swap and position-swap perturbations without relying on memorized scenes.
  • G.3 Anchor-Align Real-World Rollouts: All six mug rollouts succeed despite simultaneous swaps of the mug’s position and surrounding distractor objects.Each rollout uses a unique scene configuration, testing robustness beyond memorized object arrangements.
  • G.3 Anchor-Align Real-World Rollouts: Anchor-Align VLA selects pineapple, red bell pepper, or yellow bell pepper from language instructions and places the target correctly in very cluttered scenes.Many distractors, including alternative candidate objects, require grounding the instruction in the current observation.

H Qualitative Results in LIBERO Simulator

Qualitative LIBERO examples show Anchor-Align succeeding under background, robot, object, camera, and lighting perturbations where standard BC fails. The advantage extends to long-horizon tasks and novel object arrangements by preserving task-relevant visual grounding and sequential behavior.

  • LIBERO-Goal Plus: Anchor-Align places the bowl on the plate under a stone-tile background and opens the middle drawer after the robot’s initial configuration shifts, unlike standard BC.Standard BC collapses onto the table in the bowl task and approaches the wrong cabinet side in the drawer task.
  • LIBERO-Goal Plus: Under rearranged objects and a changed floor texture, Anchor-Align locates and opens the middle drawer and places the bowl on the cabinet, while standard BC misses the relevant regions.The object perturbation rearranges the cabinet, plate, and distractors; the background perturbation swaps the wood floor for stone tile.
  • LIBERO-Long Plus: Anchor-Align sequentially handles both targets in long-horizon basket tasks under shifted viewpoints, whereas standard BC selects an incorrect object.The camera perturbation changes the viewpoint in the task involving the cream cheese box and butter.
  • LIBERO-Long Plus: Anchor-Align deposits both required items under darkened lighting and completes the black-bowl drawer task despite camera perturbation, while standard BC grasps incorrectly or idles.The lighting example involves alphabet soup and tomato sauce; the camera example requires placing the black bowl in the bottom drawer and closing it.
  • LIBERO-PRO: On LIBERO-PRO object swaps with novel surrounding objects, Anchor-Align grounds to the alphabet-soup can and places it in the basket, unlike standard BC.The basket is replaced with a pink wicker variant and surrounding canned items are exchanged for unseen objects.
Loading 2607.13429v1…