💻 computer science

Revisiting Embodied Chain-of-Thought for Generalizable Robot Manipulation

This paper introduces ERVLA, a vision-language-action model that leverages a large-scale embodied chain-of-thought corpus for representation-shaping supervision rather than autoregressive inference, achieving state-of-the-art generalization and stability in robot manipulation tasks.

Original authors: Nan Sun, Yuan Zhang, Yongkun Yang, Wentao Zhao, Peiyan Li, Jun Guo, Wenxuan Song, Pengxiang Ding, Runze Suo, Yifei Su, Xin Xiao, Xinghang Li, Huaping Liu

Published 2026-06-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nan Sun, Yuan Zhang, Yongkun Yang, Wentao Zhao, Peiyan Li, Jun Guo, Wenxuan Song, Pengxiang Ding, Runze Suo, Yifei Su, Xin Xiao, Xinghang Li, Huaping Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to do chores, like putting a toy car into a drawer. In the past, scientists tried to teach robots by giving them a "brain" (a Vision-Language Model) that could understand pictures and words, and a "hand" (an Action Model) that could move. But often, the brain would get confused, and the hand would fumble.

This paper introduces a new way to teach robots called ERVLA (Embodied Reasoning Vision-Language-Action). Here is the story of how they did it, explained simply.

1. The Problem: The Robot That Talks Too Much (and Too Slowly)

Imagine you are driving a car, and your GPS tells you every single turn before you make it. "In 100 feet, turn left. Now, turn left. Now, turn left." If the GPS makes a tiny mistake in the first sentence, the rest of the directions get messed up, and you crash.

This is what happened with previous robot "thinking" methods (called Embodied Chain-of-Thought). The robot was forced to "think out loud" (write a long text plan) before it could move its arm.

The Issue: If the robot wrote a slightly wrong sentence in its plan, the rest of the plan would fail. It was slow, and one small error would ruin the whole task.
The Paper's Discovery: They found that making the robot "talk more" doesn't make it smarter. In fact, forcing it to write a long plan before moving often makes it worse.

2. The Solution: "Thinking" While Doing, Not Before

The authors realized that the robot doesn't need to say its thoughts out loud to be smart. It just needs to have the thoughts inside its brain while it moves.

Think of it like a master chef. A novice chef might write a recipe step-by-step on a piece of paper before cooking. A master chef doesn't write anything down; they just know what to do because they have practiced the steps so many times. Their "thinking" is built into their muscle memory.

ERVLA teaches the robot to be like the master chef:

The Training: They gave the robot a massive library of 978,000 robot movements (like a giant cookbook).
The Trick: They taught the robot to read the "recipe" (the plan) and the "action" (the movement) at the same time.
The Secret Sauce (Reasoning Dropout): Sometimes, during training, they told the robot, "Don't write the recipe this time, just move!" This forced the robot to learn how to understand the idea of the task without needing to write it down first. It learned to internalize the reasoning.

3. The "CoT Contamination" Problem

The paper also found a hidden trap. When they used computers to automatically write these "recipes" for the robot, sometimes the computers made mistakes (like saying the cup is in the wrong spot).

If the robot tried to memorize these wrong recipes, it got confused. This is called CoT Contamination.
The Fix: They taught the robot to ignore the parts of the recipe that looked shaky or unreliable, focusing only on the clear, solid instructions.

4. The Results: A Robot That Actually Works

They tested this new robot (ERVLA) in two ways:

In the Simulator (Video Game): It became the best robot in the world at its specific tests, beating all previous models. It could handle tricky situations where the lighting changed or objects were moved.
In the Real World: They put it on a real physical robot arm.
- When asked to "put away the thing that isn't a fruit" (a tricky, vague instruction), other robots got confused. ERVLA understood the logic and did it.
- When asked to do a long, multi-step task (like clearing a whole table), ERVLA kept its cool and finished the job, while others gave up or got lost.

Summary Analogy

Old Way: The robot is a student who must write a 10-page essay before it is allowed to take a single step. If it misspells a word in the essay, it fails the test.
ERVLA Way: The robot is an experienced athlete. It has practiced so much that it understands the game plan instantly. It doesn't need to write an essay to know how to play; the strategy is built into its movements.

The Bottom Line: The paper shows that for robots to be smart, they shouldn't be forced to "talk" their way through every task. Instead, they should be trained to absorb the logic of the task so that their actions naturally follow the right path, even when the instructions are vague or the world is messy. They also released the massive "cookbook" (dataset) and the robot's "brain" (model) for others to use.

Technical Summary: Revisiting Embodied Chain-of-Thought for Generalizable Robot Manipulation

Problem Statement

Vision-Language-Action (VLA) models have emerged as a promising paradigm for generalizable robot control by transferring semantic priors from large-scale Vision-Language Models (VLMs) into embodied decision-making. However, enhanced perception and broader semantic coverage do not inherently guarantee better action generation. VLA models often struggle in complex, out-of-distribution scenarios that require deliberation prior to execution.

While previous Embodied Chain-of-Thought (ECoT) methods have attempted to bridge this gap by supervising explicit reasoning traces, several critical issues remain unresolved:

Form and Integration: It is unclear which forms of embodied reasoning (e.g., scene understanding, spatial grounding, trajectory prediction) genuinely facilitate effective action generation versus those that merely add verbal overhead.
Interaction with Policy: Early ECoT formulations model reasoning as an autoregressive prefix to action generation. This design is brittle; long reasoning traces increase latency, and errors in the reasoning prefix compound during inference, negatively affecting subsequent action tokens.
Scalability: The scaling behavior of CoT data is unclear. Large-scale embodied CoT datasets are scarce due to the high cost of annotation, and it is unknown whether scaling reasoning supervision yields stronger action generation or introduces noise that hinders learning.
CoT Contamination: In large-scale auto-labeled datasets, noisy labels (e.g., unstable bounding boxes, drifting end-effector coordinates) can impose inconsistent supervision on semantically similar observations, hindering adaptation.

Methodology: ERVLA

The authors propose ERVLA (Embodied Reasoning Vision-Language-Action), a model designed to effectively leverage linguistic reasoning for generalizable robot manipulation without relying on mandatory test-time CoT.

1. Large-Scale Embodied CoT Corpus

The authors curated the largest embodied CoT corpus to date, comprising:

978,743 trajectories
226.3 million samples
2,592.5 hours of data
Sources include Bridge, Fractal, Droid, MolmoAct, and AgiBot datasets.

The data is structured into a hierarchical schema decomposing reasoning into four categories:

Understanding: Task goals and intent.
Grounding: Spatial alignment with visual entities (e.g., object bounding boxes, gripper positions).
Planning: Subtask decomposition and progress tracking.
Acting: Action-oriented motion descriptions and future point trajectories.

2. Model Architecture

ERVLA integrates a Vision-Language Reasoning module with a Diffusion-based Action module:

Backbone: A VLM (specifically Qwen3-VL-4B) processes visual observations and text instructions. Unlike "ThinkAct" styles that separate planning and action, ERVLA injects action-sensitive supervision directly into the VLM.
Action Head: A Diffusion Transformer (DiT) generates continuous robot actions via flow matching.
Knowledge Truncation: The DiT conditions only on the key-value (KV) cache of the semantic prefix (reasoning) from the VLM, excluding appended control-query tokens. This prevents the action model from exploiting synthetic shortcuts and reduces interference from training-specific token patterns.
Auxiliary Action Queries: The VLM is trained with an auxiliary action-regression loss through query tokens. This guides the backbone to understand action-relevant semantics, not just produce text.

3. Training Strategy: CoT-Dropout

A central innovation is the CoT-dropout strategy. During training, the model is exposed to samples with explicit CoT supervision and samples without it (where the CoT span is empty) with a certain probability ( $p_{cot}$ ).

Purpose: This allows the model to leverage rich reasoning traces during training to reshape its internal representations but enables it to predict actions directly without CoT during inference.
Benefit: This bypasses the autoregressive instability of CoT prefixes, mitigates "CoT contamination" from noisy labels, and treats reasoning as a training signal that reshapes the representation space rather than a mandatory verbal output.

Key Contributions

Dataset Construction: Creation of the largest and most comprehensive embodied CoT corpus, featuring multi-view and bimanual settings with structured, hierarchical annotations.
Empirical Analysis of CoT Signals: Through extensive ablation studies, the authors identify that:
- High-level semantic understanding alone yields marginal gains.
- Effective CoT must ground high-level semantics in concrete linguistic action guidance (e.g., end-effector movement, image-space trajectories).
- Explicit CoT does not scale reliably as an autoregressive action prefix due to compounding errors.
- Noisy grounding fields (e.g., unstable bounding boxes) can be harmful unless regularized via reasoning dropout.
ERVLA Model: A novel VLA architecture that internalizes reasoning into action-aware representations. It achieves state-of-the-art performance by combining explicit CoT supervision, reasoning dropout, auxiliary action queries, and knowledge-truncated conditioning.
Scaling Insights: The paper demonstrates that embodied CoT scales effectively when reasoning supervision and action losses are jointly optimized in the representation space, rather than treated as literal autoregressive targets.

Experimental Results

The authors evaluated ERVLA on simulation benchmarks and real-robot platforms.

Simulation Benchmarks

LIBERO-Plus: ERVLA achieved a state-of-the-art success rate of 86.9%, surpassing competitive baselines like $\pi0.5$ (85.5%). It demonstrated strong zero-shot generalization, reaching 100% success on background and lighting variations in the Spatial track.
VLABench: ERVLA achieved an average success rate of 53.2%, outperforming strong baselines in challenging out-of-distribution settings requiring semantic understanding and instruction following.
Scaling: Experiments showed that ERVLA benefits from increased embodied CoT pre-training data, whereas autoregressive CoT+action-token models degrade as data scales due to noise accumulation.

Real-Robot Experiments

Deployed on physical robots with third-person and wrist cameras, ERVLA was tested on tasks involving drawer placement and table clearing across four difficulty tiers: Basic, Distractors, Semantic, and Long-horizon.

Performance: ERVLA outperformed baselines (ECoT, WorldVLA, UniVLA, $\pi0.5$ ), particularly in handling semantic ambiguity (e.g., "put away the object that is not a fruit") and long-horizon tasks.
Robustness: While baselines struggled with autoregressive latency and error propagation, ERVLA's ability to internalize reasoning allowed for more robust execution under semantic shifts and distractors.

Significance and Claims

The paper argues that embodied CoT should be understood less as an output format (a robot "saying more") and more as an interface between semantic reasoning and action learning.

Core Thesis: Reasoning benefits robot policies not by being longer or more explicit, but when its supervision is reliable, action-relevant, and integrated into the policy in a way that shapes representations rather than burdening decoding.
Failure Modes: The work highlights that naive autoregressive coupling of CoT and action leads to scaling failures and that noisy grounding signals can contaminate learning if not regularized.
Future Direction: The authors position ERVLA as a foundation for future work on reasoning-aware robot learning, emphasizing the need for reliable grounding supervision, adaptive test-time reasoning, and architectures that treat reasoning as a learnable bridge from perception to action rather than verbal decoration.

The paper concludes that the key to generalizable robot manipulation lies in converting pretrained semantic priors into action-relevant representations through structured, internalized reasoning, rather than relying on explicit verbal traces during execution.