Explicit Action Reasoner (EAR) in Robotic Policies

• Explicit Action Reasoner (EAR) is a neural sequence model that produces coarse-grained action trajectories from multimodal observations, improving the mapping between perception and action.
• It integrates into ACoT-VLA pipelines using an 18-layer transformer with self- and cross-attention to generate explicit guidance embeddings for downstream action prediction.
• Empirical results on LIBERO benchmarks demonstrate that EAR significantly improves manipulation success rates, validating its effectiveness in action-centric deliberative reasoning.

The Explicit Action Reasoner (EAR) is a neural sequence model component proposed for vision-language-action (VLA) architectures under the Action Chain-of-Thought (ACoT) paradigm. Unlike prior reasoning intermediates—such as sub-task language or visual goal representations—EAR produces explicit, coarse-grained action trajectories, allowing the reasoning process to directly unfold in action space. EAR thus serves as an action-level inductive bias, enabling grounded and efficient translation of multimodal observations into physically executable controls. Its integration in ACoT-VLA demonstrably elevates manipulation success rates on major robot learning benchmarks, supporting its role as a core operator for action-centric deliberative reasoning in robotic policy architectures (Zhong et al., 16 Jan 2026).

1. Conceptual Role of EAR in ACoT

Within ACoT, explicit action reasoning is positioned as a more effective approach compared to conventional intermediate reasoning in language or vision domains. EAR computes coarse reference trajectories—referred to as $g_{action}^{ex}$—directly from the multimodal context, thereby decreasing ambiguity in the perception-to-action mapping and serving as an intermediate “thought process” in kinematic space. These trajectories inform and guide the downstream action head, improving policy reliability and robustness in complex robotic manipulation tasks.

2. Architectural Integration and Data Flow

EAR operates as part of the ACoT-VLA pipeline, which consists of a shared VLM backbone (SigLIP visual encoder and Gemma-2B LLM, $N=18$ layers, $d=2048$), the EAR and its complementary Implicit Action Reasoner (IAR), and an Action-Guided Prediction (AGP) head. The sequence is as follows:

• The VLM backbone produces multimodal keys ($K^{VLM}_i$) and values ($V^{VLM}_i$) for each transformer layer $i=1\dots N$ from the observation $o_t$ and natural language command $l$.
• EAR receives the cached VLM features and a noisy action chunk $$\tilde a_{t:t+H^{ref}-1} \in \mathbb{R}^{H^{ref}\times A}$$ as input, denoises it, and produces a coarse action trajectory $a^{ref}_{t:t+H^{ref}-1}$.
• This trajectory is projected via an MLP to construct an explicit guidance embedding $Z^{ex}$ for use by the action prediction head.

3. Neural Module Structure and Mathematical Formulation

EAR is an 18-layer transformer sequence model characterized by self- and cross-attention mechanisms at each layer. The internal forward flow at layer $i$ comprises:

• Embedding: $h_0^{ref} = \mathrm{Embed}(\tilde a)$
• Self-Attention: $S = \mathrm{SelfAttn}(h_{i-1}^{ref})$
• Cross-Attention: $$C = \mathrm{CrossAttn}(h_{i-1}^{ref}, K^{VLM}_i, V^{VLM}_i)$$
• Feedforward/Residual: $h_i^{ref} = h_{i-1}^{ref} + \mathrm{FFN}(S + C)$

The top-layer hidden $h_N^{ref}$ is linearly decoded to produce the denoised action trajectory $a^{ref}_{t:t+H^{ref}-1}$, which is subsequently projected as $Z^{ex} = \mathrm{MLP}(a^{ref}_{t:t+H^{ref}-1})$.

The loss for EAR is a flow-matching mean-squared error (MSE):

$$\mathcal{L}_{\pi^{ref}_\theta} = \mathbb{E}_{t,\text{noise}} \left[ \| \mathrm{DenoisePred} - \text{target} \|^2 \right]$$

The total training loss combines this with the action head loss:

$$\mathcal{L}_{\mathrm{total}} = \lambda_1 \mathcal{L}_{\pi^{ref}_\theta} + \lambda_2 \mathcal{L}_{\pi^{head}_\theta}, \qquad \lambda_1=\lambda_2=0.5$$

4. Algorithmic Workflow and Integration with Action-Guided Prediction

The operational flow of EAR during training and inference is as follows:

1. Encoding: Compute VLM caches $(K_i, V_i)$ for each layer from the observation and language.
2. Input Noising: Generate noisy reference actions via self-conditioning: $\tilde a = a_{gt} * \alpha + \epsilon * (1−\alpha)$, with $\epsilon \sim \mathcal{N}(0,I)$.
3. EAR Processing: Embed $\tilde a$, then apply $N$ blocks of self-attention, cross-attention to VLM caches, and parallel FFNs to yield $h^N_{ref}$.
4. Decoding and Projection: Produce $a_{ref}$ and form the explicit guidance vector $Z^{ex}$ using an MLP.
5. Downstream Prediction: AGP head consumes $Z^{ex}$ (explicit) and IAR output (implicit), fusing both in a transformer block to generate the final action prediction.

During inference, EAR uses its own predicted $a^{ref}$ instead of teacher-forced ground truth to construct $Z^{ex}$.

5. Training Procedures and Hyperparameterization

EAR is trained using demonstration datasets from LIBERO (1,693 episodes), LIBERO-Plus (14,347 episodes), and VLABench (4,713 episodes), employing standard data splits and no external data. Action representations include delta end-effector (Delta-EEF, $H^{ref}=15$, shift=2) for simulation and absolute EEF or joint commands for real-world tasks. Loss balancing ensures stable training by maintaining $\lambda_1 = \lambda_2 = 0.5$. Teacher-forcing is used during training: $Z^{ex}$ is derived from ground-truth $a_{gt}$ to prevent destabilization of the action head.

6. Empirical Results and Scalability Analysis

Empirical ablations on LIBERO and LIBERO-Plus benchmarks demonstrate the efficacy of EAR:

Dataset	Baseline SR	+EAR Only	+IAR Only	+EAR+IAR
LIBERO	96.9%	98.3%	98.1%	98.5%
LIBERO-Plus	75.7%	83.7%	80.4%	84.1%

The introduction of EAR yields substantial improvements in average success rate (SR), outperforming baselines and configurations with only implicit reasoning, especially in more complex task distributions. EAR exhibits robust performance across reference-action hyperparameters (e.g., $H^{ref}=15$, shift=2), and a parameterized scaling analysis (Table 13) indicates that moderate EAR size ($\sim$300M parameters) provides optimal gains, while further increases can induce overfitting. Qualitative results show that EAR’s coarse trajectory “thought tokens” enhance physical manipulation reliability in both simulated and real environments (Zhong et al., 16 Jan 2026).

7. Implications and Technical Significance

The Explicit Action Reasoner operationalizes the ACoT paradigm’s core hypothesis: explicit, structured reasoning in action space confers inductive bias and improves the mapping from multimodal input to grounded robotic actions, particularly for generalist policies tasked with diverse manipulation. Its transformer-based architecture, explicit guidance mechanism, and favorable scaling properties collectively underpin its technical contribution to robot policy learning. The experimental results establish EAR as an essential operator for explicit action-space deliberation, setting a precedent for future architectures in vision-language-action modeling.