RoboFollow: Unveiling the Instruction Following Mirage in Embodied Agents

AI summary

Overview

Research area: Embodied artificial intelligence, specifically Vision-Language-Action (VLA) models and World Action Models (WAM), plus benchmark design for robotic manipulation and instruction following.

Technical level: Intermediate. The paper's argument is conceptually simple, but readers benefit from some familiarity with language-conditioned robot policies, VLA/WAM architectures, and manipulation benchmarks such as LIBERO and RoboTwin2.0.

One-sentence scope: The paper introduces RoboFollow, a diagnostic benchmark built on high-entropy scenes where the visual layout alone cannot identify the task, and uses a four-level evaluation protocol (L0–L3) with stage-wise Intent and Execution scores to test whether nine VLA and WAM policies genuinely follow language instructions.

What This Paper Is About

Modern embodied agents report impressive task success rates, but the authors argue these numbers hide a weaker reality: the policies often succeed without really using the language instruction, because the visual scene already implies a single plausible action. RoboFollow is designed to remove that shortcut by pairing each scene with multiple kinematically distinct, semantically valid task branches, so that a policy must read the instruction to pick the right behavior. The goal is to diagnose whether instruction following is genuine, and where it breaks down, rather than to measure manipulation competence in general.

Key Contributions

1. A language-necessary diagnostic benchmark with controlled execution confounds. RoboFollow builds high-entropy scenes in which one visual configuration supports several feasible task branches, and deliberately uses simple objects, short-horizon interactions, and action primitives covered by training to reduce motor-execution confounds.
2. A hierarchical diagnostic protocol with stage-wise scoring. The L0–L3 protocol progressively tests in-distribution execution, visual grounding under layout changes, semantic recombination under familiar layouts, and joint visual–semantic generalization; Intent and Execution Scores separate errors in intent selection from execution inaccuracies given a correct intent, and a dedicated finishing stage accounts for 20% of the score.
3. A systematic analysis of current models and mitigations. Nine state-of-the-art VLA and WAM policies are evaluated, alongside stronger VLM backbones, QA co-training, LangForce, and classifier-free guidance, showing that none of these strategies closes the generalization gap beyond L0.

Main Findings

• High scene entropy is the design lever. RoboFollow contains 75 training task labels grouped into six task-independent initial scene configurations, with K_s = (16, 16, 16, 16, 7, 4) and K = 75, yielding 3.782 bits of scene entropy, compared with 0.880 bits for the equally weighted LIBERO Spatial/Object/Goal/Long suites under the same task-label definition.

• Strong in-distribution scores do not transfer. Evaluated across nine VLA and WAM policies, all models show a pronounced decline in Intent Score from L0 to L1–L3, and the consistent drop in Intent Score indicates the degradation cannot be explained by low-level execution failure alone.

• Sharp concrete collapses. On Scenes 1 and 2, the π-series models reach near-saturated L0 performance: π0.5 obtains 99.1% and 100.0% Intent Score, and π0 reaches 98.2% and 95.6%. In Scene 1, π0 falls from 98.2% at L0 to 0.0% at L1, while π0.5 declines from 99.1% to 45.5% at L1 and 44.9% at L3. In Scene 2, π0.5 decreases from 100.0% at L0 to 34.2% at L3.

• Language is doing relatively little work. A language-ablated variant of PI05 ("empty language") reports average Intent Score 12.3, Execution Score 8.9, and Completion Rate 5.2 across all scenes and levels, with per-scene L0 Intent Scores of 16.0 (S1), 16.6 (S2), 6.4 (S3), and 20.4 (S4), suggesting the model can often still make progress without an instruction.

• The VLM backbones themselves are weak at scene understanding. On a 20-question visual QA probe targeting object identities, colors, and spatial relations in Scene 2, accuracy was 1/20 for base PaliGemma, 2/20 for the pre-trained π0.5, and 3/20 for the fine-tuned π0.5 backbone.

• There is a two-layered failure structure. Even for questions the fine-tuned VLM answers correctly, about half of those episodes still end in incorrect manipulation behavior, indicating a comprehension-to-execution gap. A comparable disconnect appears in Motus, whose VLM backbone is entirely frozen during training.

• A much stronger VLM does not fix it. Qwen3-VL (4B) achieved 19/20 on the Scene 2 QA probe, yet a Qwen-GR00T model combining that backbone with the GR00T diffusion action head and QA co-training still showed a catastrophic collapse in out-of-distribution instruction following across L1–L3.

• Existing mitigations fall short. LangForce yields only marginal changes, which the authors attribute to amplifying text–action statistical correlation without supervising compositional semantic understanding. Classifier-Free Guidance is counterproductive: raising the guidance scale to 1.2 and 1.5 degrades even L0 (pi05_cfg_1.2 reaches 55.9% L0 Intent Score and pi05_cfg_1.5 reaches 43.6%, versus 93.6% for the pi05 baseline on Scene 2) and produces erratic trajectories.

• Fine-tuning controls help but leave large gaps. On the 16 Scene 2 tasks with π0.5, doubling demonstrations from 25 to 50 per task improves all Intent Scores but changes the L0–L2 gap only from 54.9 to 55.8 percentage points and the L0–L3 gap from 48.6 to 47.8. Increasing instruction variants from one to five improves L1 (72.8% to 79.0%) and L3 (41.5% to 53.0%) while L2 varies non-monotonically (31.6%, 37.8%, 32.0%). From 2k to 6k training steps, L0/L2 improve and L1 decreases, with L3 peaking at 4k; at 6k the L0–L2/L3 gaps remain 55.8/47.8 points.

• A small real-robot pilot shows the same direction. With π0.5, using eight training instructions and eight held-out instructions over the same object set and testing each instruction five times, success fell from 20/40 (50%) on training instructions to 6/40 (15%) on held-out instructions. The authors note this preliminary comparison uses different instruction sets and pick/stack compositions rather than matched task pairs.

Methodology in Plain English

The researchers started from a diagnosis of why success rates are misleading: in most benchmark episodes, the initial camera view already implies one sensible action, so the instruction is partly redundant. They therefore designed scenes where this is not true. In each scene, the same or nearly the same object layout can support several different valid tasks, so a policy has to read the language to know what is wanted. They formalized this ambiguity as scene entropy, the conditional entropy of training task labels given the task-independent scene specification, and used it as a design rule rather than a scoring metric.

The benchmark has four scene families: extrinsic spatial relations (for example "left of", "right of", "behind"), intrinsic object properties combined with actions such as pick, push, stack, and place, fine-grained action modulation with waypoints and final orientations, and elementary logical grounding including temporal sequencing, negation, and conditional branching. Each family is evaluated at four levels. L0 uses the training distribution. L1 keeps the instruction fixed but swaps object positions, testing whether language binds to the right physical entity rather than memorized coordinates. L2 keeps the layout fixed but recombines semantic attributes seen in training, testing compositionality. L3 changes both.

Scoring is split into two parts. The Intent Score checks whether the policy selected the right object, relation, waypoint, action primitive, orientation, or logical branch at each stage. The Execution Score checks whether that subgoal was physically completed. A Completion Rate provides an action-dependent binary criterion as a complement. This design lets the authors distinguish semantic misunderstanding from low-level control failure and penalizes policies that reach the final state through wrong intermediate actions.

The benchmark is built on the RoboTwin2.0 simulation platform, with 3,750 training episodes in total and 50 episodes per task; when object arrangements repeat across episodes, a slight random positional perturbation of 1–2 cm is applied, and each training instruction is randomly sampled from paraphrased variants. Main experiments fine-tuned on all the data; the analysis section used only 16 tasks (800 episodes) from Scene 2 for fine-tuning. Beyond the main comparison, the authors ran a visual QA probe, trained a Qwen-GR00T variant, applied LangForce, applied classifier-free guidance at two scales, and swept demonstration count, instruction variants, and training duration.

Why This Matters

Impact on research. The paper reframes instruction following as a measurement problem: if benchmark scenes have low entropy, high success rates can coexist with policies that barely use language. The scene entropy formulation (3.782 bits for RoboFollow versus 0.880 bits for LIBERO Spatial/Object/Goal/Long) gives the community a quantitative design principle for building more diagnostic datasets, and the Intent/Execution split gives a way to attribute failures instead of lumping them into a single success number. The negative results for stronger VLM backbones, QA co-training, LangForce, and classifier-free guidance also set a clear baseline that future methods must beat.

Real-world applications (implied by the paper's framing, not reported as deployed results):

• Warehouse or factory robots that must choose among several visually similar objects based on a spoken or written order.
• Home assistive robots where the same tabletop arrangement could call for a pick, a push, or a stack, and only the instruction disambiguates.
• Surgical or laboratory automation where procedural constraints such as waypoints and final orientations matter, not just reaching an end state.
• Safety-relevant settings where a visually successful action that ignores the instruction is semantically wrong and potentially harmful.

Industry relevance. Teams deploying language-conditioned manipulation policies need to know whether their systems are genuinely controllable through language. The paper's results suggest that a policy can score near 100% Intent Score in distribution and still collapse to 0.0% under a simple object swap, which is exactly the kind of failure that is hard to detect in standard acceptance testing. The finding that classifier-free guidance degrades performance in high-entropy scenes is directly actionable for practitioners tuning diffusion-based policies.

Future Directions

• Increase structural layout diversity during training. The authors explicitly note this remains untested, because evaluating it requires new training layouts while keeping evaluation layouts held out; training on existing L1 test layouts would invalidate the split.
• Supervise grounded language-to-action alignment during adaptation. The authors propose this, plus increasing structural scene and semantic diversity while preserving held-out combinations, as hypotheses rather than demonstrated fixes.
• Close the comprehension-to-execution gap. Since a backbone answering 19/20 QA questions (Qwen3-VL-4B) still failed at L1–L3 after integration, better perception alone is not enough; the mechanism by which understood semantics fail to reach the action head is unresolved.
• Extend beyond the controlled setting. The limitations section calls for extending RoboFollow to richer embodiments and real-world scenarios while preserving its language-necessary design, and notes that long-horizon planning, contact-rich manipulation, open-vocabulary diversity, and large-scale real-world deployment are not covered. Transfer across simulators and to more complex real-world tasks has not been established.

Target Audience

Researchers and engineers working on embodied AI, VLA and WAM models, and robot learning evaluation. It is most useful for people who build or benchmark language-conditioned manipulation policies, for teams deciding which training or inference tricks to invest in (since several popular ones are shown to be insufficient), and for benchmark designers interested in how scene construction shapes what a dataset can actually diagnose. Readers without a robotics background can follow the core argument, but the experimental tables assume familiarity with metrics and model families.

Authors’ abstract

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.