PaperlayerView exact revision on arXiv

Source-linked AI summary

RoboFollow: Unveiling the Instruction Following Mirage in Embodied Agents

Chang Guo, Yukun Xie, Bohan Tan, Zheng Chang, Zhaokai Yin, Qianli Ma, Yingqiao Wang, Chao Liang, Zhipeng Zhang

arXiv:2609.25636v1cs.RO

TL;DR

High task success can mask weak instruction following when visual scenes make language partly redundant. RoboFollow builds a benchmark with ambiguous task scenes, hierarchical tests, and separate intent and execution scores; evaluated policies’ strong L0 performance does not reliably transfer to L1–L3 under the fine-tuning setup.

  • Problem

    Manipulation success measures can entangle visual recognition, language grounding, planning, and control, while visually dominant behaviors can make language partly redundant.

    Source: §1 · p. 1

  • Method

    RoboFollow creates scenes with multiple feasible task branches and combines L0–L3 evaluation with stage-wise Intent and Execution scoring.

    Source: §1 · p. 2

  • Results

    Across evaluated VLA and WAM policies, strong L0 performance does not reliably transfer to L1–L3, and tested optimizations do not close the gap.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Robust instruction following remains a critical bottleneck for controllable and reliable embodied agents.

    Source: §6 · p. 9

    RoboFollow simplifies object geometry, task horizons, and interaction dynamics, and does not fully cover long-horizon planning, contact-rich manipulation, open-vocabulary diversity, or large-scale real-world deployment.

    Source: §7 · p. 9

Abstract

from arXiv · show

Modern embodied agents achieve impressive success rates, yet their actual instruction-following ability is far weaker than these numbers suggest. We trace this illusion to a structural property we term low scene entropy: when a visual scene admits only one valid task, language becomes redundant and a policy can score highly while barely using it. We introduce RoboFollow, a diagnostic benchmark with three principles: (1) High Scene Entropy: each training scene supports multiple kinematically distinct task branches, making vision alone insufficient and forcing reliance on language. (2) Hierarchical Diagnostic Protocol: a four-level protocol (L0--L3) progressively perturbs visual layout and semantics, probing whether equivalent instructions yield consistent behavior and distinct ones yield discriminable behavior across spatial relations, attributes, trajectory constraints, and logic. (3) Confound-Controlled Diagnosis: we simplify interaction objects, restrict actions to the trained repertoire and report stage-wise Intent and Execution scores, isolating comprehension from motor execution. Evaluation of nine VLA and WAM policies shows that strong L0 performance, where attained, does not reliably transfer to L1--L3 under our fine-tuning setup. Representative mitigations, including stronger VLM backbones, QA co-training, LangForce, and Classifier-Free Guidance, all fail to close this gap. RoboFollow exposes genuine instruction following as a critical, overlooked bottleneck. Code and dataset are available at https://github.com/AutoLab-SAI-SJTU/RoboFollow and https://huggingface.co/datasets/AutoLab-SJTU/robofollow-data.

1 Introduction

High task success can mask weak instruction following when the scene already suggests a plausible action and makes language partly redundant. RoboFollow creates ambiguity to test language-dependent task selection, separates intent from execution, and finds that strong in-distribution performance does not reliably transfer under the evaluated fine-tuning setup.

1 Introduction

  • High task success may reflect a visually suggested action rather than instruction use, even though semantically wrong actions matter for deployable robots.Standard manipulation success measures entangle visual recognition, language grounding, planning, and low-level control, while canonical behaviors can make language redundant.Source: §1 · pp. 1–2
  • RoboFollow makes language necessary by placing multiple semantically valid, kinematically feasible task branches in the same or highly similar visual scene.It quantifies this ambiguity as scene entropy: the conditional entropy of training task labels given the task-independent initial scene specification.Source: §1 · p. 2
  • The benchmark combines four task families, L0–L3 testing, and stage-wise Intent and Execution Scores to distinguish semantic misunderstanding from motor failure.The families cover spatial relations, object attributes and action selection, trajectory and orientation constraints, and elementary logic.Source: §1 · p. 2
  • Across evaluated VLA and WAM models, strong L0 performance does not reliably transfer: Intent Scores degrade sharply under L1–L3, and tested optimizations do not close the gap.The evaluation includes stronger VLM backbones, QA co-training, LangForce, and classifier-free guidance.Source: §1 · p. 2
  • RoboFollow uses simple objects, short-horizon interactions, and trained action primitives to reduce motor-execution confounds while evaluating instruction following.Its analysis evaluates representative VLA and WAM models alongside stronger VLM backbones, QA co-training, LangForce, and classifier-free guidance.Source: §1 · p. 2

2 Related Work

Prior work spans VLA and WAM architectures for connecting perception, language, and action, alongside manipulation benchmarks that test varied capabilities and robustness. RoboFollow addresses training-time task ambiguity by pairing shared scenes with multiple task specifications.

2.1 Vision-Language-Action and World Action Models

  • VLA methods connect pretrained vision-language models to physical execution, while WAMs couple action generation with predictive modeling of environment dynamics.The cited VLA approaches include autoregressive action generation, flow matching, grounded representations, quantization, and visual chain-of-thought reasoning.Source: §2.1 · pp. 2–3

2.2 Benchmarks for Robotic Manipulation Evaluation

  • Manipulation benchmarks cover standardized control, long-horizon interaction, multimodal reasoning, semantic generalization, scalable demonstrations, and robustness; LIBERO variants report visual-pattern exploitation over language semantics.RoboFollow pairs shared training scenes with multiple task specifications, complementing LIBERO-PRO’s test-time perturbations; π0.7 and IVA address different settings.Source: §2.2 · p. 3

3 RoboFollow Benchmark

RoboFollow creates shared visual scenes that support multiple feasible task branches, then probes language grounding through varied semantic scenes, visual-semantic test levels, and stage-wise scores. Scene entropy characterizes training ambiguity, while Intent and Execution scoring separates task selection from physical completion.

3 RoboFollow Benchmark

  • RoboFollow makes the initial observation insufficient to infer the target behavior by placing multiple valid, kinematically feasible task branches in the same or similar visual configuration.The policy must condition on the instruction to resolve the task ambiguity.Source: §3 · p. 3
  • The benchmark groups 75 training task labels into six task-independent initial scene configurations and defines scene entropy over task labels conditioned on those configurations.Scene entropy is used as a design principle to remove visual shortcuts, not as a final evaluation metric.Source: §3 · p. 3

3.1 Scene Design

  • The L0–L3 protocol tests in-distribution behavior, changed layouts, novel semantic recombinations, and combined visual-semantic perturbations; its levels isolate dimensions rather than rank difficulty.Training and validation splits preserve linguistic clarity, avoid leakage across levels, and keep test-time target actions within the demonstrated repertoire.Source: §3.1 · pp. 3–4
  • RoboFollow’s four diagnostic scenes test spatial relations, intrinsic object properties, action and trajectory constraints, and elementary logic.The logic scene uses short controlled instructions for sequencing, negation, and observable conditional branching, with scene state varied across splits.Source: Figure 1 · p. 4, §3.1 · pp. 4–5

3.2 Metrics

  • Multi-Stage Intent-Execution Scoring addresses the limits of binary final-state success by separately scoring correct semantic choices and completion of their corresponding subgoals.Intent covers the target at each stage; Execution covers physical completion, and incorrect intermediate actions are penalized even if the final state is reached.Source: §3.2 · p. 5
  • Completion Rate complements Intent and Execution Scores with an action-dependent binary completion criterion, averaged unweighted across tasks within each scene and level.It does not require every intermediate stage to receive full credit.Source: §3.2 · p. 5

4 Experiments and Results

RoboFollow evaluates nine VLA and WAM policies across four scenes, with the main experiments fine-tuned on all 3,750 training episodes. Strong in-distribution Intent Scores often fall sharply under visual or semantic perturbations, with differences across scene types and policies.

4.1 Experimental Setup

  • The evaluation covers nine foundation models from VLA and WAM paradigms; the four scenes contain 3,750 training episodes, and the main experiments use all the data for fine-tuning.The analysis experiments instead use 16 Scene 2 tasks (800 episodes).Source: §4.1 · pp. 5–6

4.2 Main Results

  • Across scenes, Intent Scores decline from L0 to L1–L3, indicating that semantic intent-selection failures contribute to the degradation beyond low-level execution failures.The evaluation reports results across all RoboFollow scenes and difficulty levels.Source: Table 1 · p. 6, §4.2 · p. 6
  • π0 falls from 98.2% IS at L0 to 0.0% at L1 in Scene 1; π0.5 drops from 99.1% to 45.5% at L1 and 44.9% at L3 there, and from 100.0% to 34.2% at L3 in Scene 2.At L0, π0.5 records 99.1% and 100.0% IS in Scenes 1 and 2, respectively; π0 records 98.2% and 95.6%.Source: §4.2 · p. 7
  • π0, GR00T N1.6, and Motus fail precipitously on Scene 1 L1, while π0 and Motus show more robust semantic grounding for Scene 2 attributes than Scene 1 spatial relations.π0.5 shows comparatively stronger spatial grounding, but still degrades under visual perturbations.Source: §4.2 · p. 7
  • A real-robot pilot with π0.5 reports 20/40 (50%) success on training instructions and 6/40 (15%) on held-out instructions, using different instruction sets and task compositions.Each of eight training and eight held-out instructions was tested five times, so this is not a matched-task comparison.Source: §4.2 · p. 7

5 Analysis

The analysis finds failures both in scene comprehension and in translating understood semantics into actions. Tested mitigation strategies and fine-tuning controls do not eliminate the generalization gap beyond L0.

5.1 VLM Deficits and The Comprehension-Execution Gap

  • Scene 2 QA accuracy is 1/20 for base PaliGemma, 2/20 for pre-trained π0.5, and 3/20 for fine-tuned π0.5; even correctly answered questions often fail to produce correct actions.About half of episodes with correct answers from the fine-tuned VLM still result in incorrect manipulation, revealing a comprehension-to-execution gap.Source: §5.1 · p. 7

5.2 Can Stronger VLMs and Existing Optimizations Help?

  • Qwen3-VL-4B scores 19/20 on Scene 2 QA, but Qwen-GR00T with QA co-training still collapses on L1–L3 instruction following.The authors conclude that a stronger VLM and preserved QA capability are insufficient to bridge the comprehension-to-execution gap in these high-entropy scenarios.Source: §5.2 · p. 8
  • LangForce yields only marginal changes, while CFG scales of 1.2 and 1.5 degrade L0 performance and produce erratic trajectories.The authors attribute LangForce's limited effect to a lack of explicit compositional semantic supervision; in high-entropy scenes, CFG's unconditional prediction leaves a noisy residual.Source: §5.2 · p. 8
  • None of the evaluated backbone, training-objective, or inference strategies yields satisfactory generalization beyond L0 in high-entropy environments.The authors leave open whether broader data distributions or alternative training procedures could improve results.Source: §5.2 · p. 8

5.3 Additional Fine-Tuning Controls

  • Doubling demonstrations from 25 to 50 changes L0–L2 from 54.9 to 55.8 percentage points and L0–L3 from 48.6 to 47.8; five instruction variants improve L1 from 72.8% to 79.0% and L3 from 41.5% to 53.0%.Across instruction variants, L2 varies non-monotonically (31.6%, 37.8%, 32.0%).Source: §5.3 · p. 9
  • Within the tested ranges, controls improve π0.5 on Scene 2 but leave substantial L0–L2/L3 gaps; structural layout diversity during training remains untested.Testing layout diversity would require new training layouts while keeping evaluation layouts held out.Source: §5.3 · p. 9

6 Conclusion

RoboFollow combines high-entropy scene design, hierarchical L0–L3 evaluation, and separate intent and execution scoring. Under the evaluated fine-tuning setup, policies degrade under visual and semantic perturbations, while tested optimization strategies remain insufficient.

6 Conclusion

  • RoboFollow diagnoses instruction following across spatial, attribute-based, procedural, and logical semantics; results identify robust instruction following as a critical bottleneck for controllable and reliable embodied agents.The benchmark combines high-entropy scenes, an L0–L3 protocol, and decoupled intent-execution scoring.Source: §6 · p. 9

7 Limitations

RoboFollow is a diagnostic benchmark, not a comprehensive test of embodied capabilities; its controlled scenes and tasks limit the settings it covers.

7 Limitations

  • Simplified geometry, task horizons, and interaction dynamics leave long-horizon planning, contact-rich manipulation, open-vocabulary diversity, and large-scale deployment outside the benchmark's coverage.The study also does not test increased structural layout diversity during training or establish transfer across simulators or to more complex real-world tasks.Source: §7 · p. 9

A Benchmark Details

RoboFollow builds training scenes with multiple feasible tasks and tests whether policies follow instructions across spatial, attribute, trajectory, and logical challenges. Its scene-entropy calculation reports 3.782 bits for RoboFollow, versus a 0.880-bit equally weighted mean for LIBERO.

A Benchmark Details

  • The four scene families test spatial relations, intrinsic object properties, trajectory constraints and orientation, and elementary logical grounding.The training set has 3,750 episodes across four scenes; paraphrased instructions are randomly sampled, and slight 1–2 cm positional perturbations are used when arrangements remain identical.Source: A Benchmark Details · pp. 15–19
  • Scene 1 tests whether spatial references follow objects after layout changes or generalize to novel combinations of familiar concepts.For example, L1 swaps the red and green balls while retaining the instruction; L2 uses “the block to the left of the red cylinder.”Source: A Benchmark Details · pp. 16–17
  • Scene 2 tests attribute and action selection, including compositional instructions such as “red cube” when its constituent concepts appeared in training.L2 keeps the default layout while changing the instruction to “red cube.”Source: A Benchmark Details · pp. 17–18
  • Scene 3 tests trajectory and orientation constraints, including equivalent route and placement instructions under linguistic reformulation.L2 changes “avoid passing over the red slab” to “pass over the green slab” and reformulates the final orientation while preserving intended behavior.Source: A Benchmark Details · p. 18
  • Scene 4 tests logical instructions by changing observable conditions and evaluating whether the policy follows the corresponding branch.L1 removes the object inside the bowl; L3 changes both layout and the instructional execution branch.Source: A Benchmark Details · p. 19

A.1 Scene Entropy Computation

  • Scene entropy groups tasks by task-independent initial scene specifications, excluding task names, language, and goals; scenes 1–3 each have one 16-label group, while scene 4 has three groups.Scene 4’s groups contain 16, 7, and 4 task labels.Source: §A.1 · p. 19
  • 3.782 bits is the reported RoboFollow scene entropy, versus LIBERO’s 0.880-bit equally weighted mean.The LIBERO Spatial/Object/Goal/Long values are 0/0/3.322/0.200 bits.Source: §A.1 · p. 19

B Metrics Details

The evaluation decomposes tasks into functional stages and reports separate Intent and Execution scores. This is designed to distinguish semantic grounding from physical completion across scene-specific objectives.

B Metrics Details

  • Tasks use up to three stages for source grounding, action-specific manipulation, and post-action outcomes; pickup-only tasks omit Stage 2.This is intended to show whether failure occurs during grounding, manipulation, or after the main action.Source: B Metrics Details · p. 20
  • Intent Score measures semantic grounding and appropriate action direction, while Execution Score measures completion of the corresponding physical sub-goal.A policy can receive partial semantic credit for approaching the correct target without completing manipulation, while an incorrect target is penalized even if the final state looks plausible.Source: B Metrics Details · p. 20
  • Stage objectives vary by scene, covering source selection, trajectory compliance, placement and orientation, and logical forms such as negation, conditionals, and temporal order.Scenes 1 and 2 also assess post-action stabilization or retraction.Source: B Metrics Details · p. 20
  • The completion rate uses action-dependent criteria and does not require full credit at every intermediate stage.Scene 3 includes positive execution credit for final placement or orientation; subsequent actions can change the evaluated final state.Source: B Metrics Details · p. 20

C Policies Training Details

Main experiments fine-tune each policy on all 3,750 training episodes, selecting training duration by model based on in-distribution convergence. The accompanying table summarizes per-policy batch sizes and training durations.

C Policies Training Details

C.1 Real-Robot Pilot Details

In the real-robot pilot, π0.5 succeeded on fewer held-out than training-instruction trials, but the instruction sets were not matched. The appendix also lists the Scene 2 VQA questions.

C.1 Real-Robot Pilot Details

  • Table 10 lists both instruction sets and their per-instruction success counts; Figure 6 shows the setup and representative right-arm executions.The pilot used eight training and eight held-out instructions, with five trials per instruction.Source: Table 10 · p. 21, §C.1 · p. 21, Figure 6 · p. 21
  • 20/40 (50%) training-instruction trials succeeded, compared with 6/40 (15%) for held-out instructions, a descriptive difference of 35 percentage points.Each instruction was tested five times; the sets differ in task composition and do not control for task identity or difficulty.Source: §C.1 · p. 21, Table 10 · p. 21
  • The training set had four pick and four stack instructions, while the held-out set had six pick and two stack instructions.Pick success was 14/20 (70%) versus 6/30 (20%); stack success was 6/20 (30%) versus 0/10 (0%).Source: §C.1 · p. 21
  • One held-out rollout qualitatively showed a grounding error: π0.5 attempted the blue cube instead of the requested red cube.This was not a quantified Intent Score.Source: §C.1 · p. 21

D QA-20 details

E Failure Cases

Figure 7 groups policy failures into intent-understood but imprecisely executed tasks and failures of semantic reasoning or instruction grounding. The policy may also continue acting after completing a task.

E Failure Cases

  • Failures divide between understood intent with insufficient execution precision and weaknesses in semantic reasoning or instruction grounding.Figure 7 presents representative failure cases.Source: E Failure Cases · p. 22, Figure 7 · p. 23
  • Semantic failures include mishandling spatial relations, object-attribute comparisons or composition, fine-grained trajectory or orientation constraints, and logical constructs.The cited logical constructs include conditionals, temporal order, and negation.Source: E Failure Cases · p. 22
  • A policy may continue taking unintended actions after the instructed task is complete.Source: E Failure Cases · p. 22

Related papers

Loading 2609.25636v1…