PaperlayerView exact revision on arXiv

Source-linked AI summary

ERA: Transforming VLMs into Embodied Agents via Embodied Prior Learning and Online Reinforcement Learning

Hanyang Chen, Mark Zhao, Rui Yang, Qinwei Ma, Ke Yang, Jiarui Yao, Kangrui Wang, Hao Bai, Zhenhailong Wang, Rui Pan, Mengchao Zhang, Jose Barreiros, Aykut Onol, ChengXiang Zhai, Heng Ji, Manling Li, Huan Zhang, Tong Zhang

arXiv:2510.12693v1cs.AI

TL;DR

Small VLMs lack embodied knowledge and skills, while large-model systems are costly to deploy. ERA combines embodied prior learning with online RL, and ERA-3B reports strong performance across high-level and low-level tasks, though evaluations are confined to simulation.

  • Problem

    Small VLMs lack embodied knowledge and robust reasoning, and a unified framework combining prior learning with online agentic RL has been missing.

    Source: §1 · p. 2

  • Method

    ERA combines trajectory-augmented, environment-anchored, and external knowledge priors with online RL designs for context management, reward shaping, and turn-level learning.

    Source: §4.1.1 · p. 5, §4.1.3 · p. 7, §4.2.1 · p. 8

  • Results

    ERA-3B surpasses prompting-based large models and 7B-scale training baselines, averaging 65.2% on EB-ALFRED and 48.3% on EB-Manipulation.

    Source: §1 · p. 2

  • Takeaways & Limitations

    ERA offers a practical and scalable recipe for developing VLM-based agents in more realistic settings.

    Source: §6 · p. 15

    All evaluations are in simulated environments, without validation on real-world systems.

    Source: A Limitation · p. 24

Abstract

from arXiv · show

Recent advances in embodied AI highlight the potential of vision language models (VLMs) as agents capable of perception, reasoning, and interaction in complex environments. However, top-performing systems rely on large-scale models that are costly to deploy, while smaller VLMs lack the necessary knowledge and skills to succeed. To bridge this gap, we present \textit{Embodied Reasoning Agent (ERA)}, a two-stage framework that integrates prior knowledge learning and online reinforcement learning (RL). The first stage, \textit{Embodied Prior Learning}, distills foundational knowledge from three types of data: (1) Trajectory-Augmented Priors, which enrich existing trajectory data with structured reasoning generated by stronger models; (2) Environment-Anchored Priors, which provide in-environment knowledge and grounding supervision; and (3) External Knowledge Priors, which transfer general knowledge from out-of-environment datasets. In the second stage, we develop an online RL pipeline that builds on these priors to further enhance agent performance. To overcome the inherent challenges in agent RL, including long horizons, sparse rewards, and training instability, we introduce three key designs: self-summarization for context management, dense reward shaping, and turn-level policy optimization. Extensive experiments on both high-level planning (EB-ALFRED) and low-level control (EB-Manipulation) tasks demonstrate that ERA-3B surpasses both prompting-based large models and previous training-based baselines. Specifically, it achieves overall improvements of 8.4\% on EB-ALFRED and 19.4\% on EB-Manipulation over GPT-4o, and exhibits strong generalization to unseen tasks. Overall, ERA offers a practical path toward scalable embodied intelligence, providing methodological insights for future embodied AI systems.

1 Introduction

Small VLMs lack the embodied knowledge and reasoning skills needed to close the gap with costly large models, while existing work has not unified prior learning with online agentic RL. ERA combines embodied priors with online RL and reports strong results across high-level planning and low-level control.

1 Introduction

  • Large models can solve increasingly complex embodied tasks with prompting but are costly to deploy, while smaller models can lag sharply, as Claude-3.5-Sonnet scores 64.0% versus Qwen2.5-VL-7B-Instruct at 4.7% on EB-ALFRED.Smaller models often lack embodied knowledge, robust reasoning, and synergy between high-level planning and low-level grounding.Source: §1 · p. 2
  • The framework first teaches small VLMs embodied priors, then refines them through online RL.The priors comprise trajectory-augmented, environment-anchored, and external knowledge sources; RL uses self-summarization, dense rewards, and turn-level policy optimization.Source: §1 · p. 2
  • ERA-3B averages 65.2% on EB-ALFRED and 48.3% on EB-Manipulation, surpassing prompting-based large models and 7B-scale training-based baselines.The two benchmarks cover high-level planning and low-level control, respectively.Source: §1 · p. 2
  • ERA studies compact-VLM post-training by unifying prior curation for supervised fine-tuning with online agentic RL, addressing both high- and low-level embodied tasks.Its related-work comparison frames the framework as unifying these stages across both task levels.Source: §1 · p. 2, Table 1 · p. 3

2 Related Work

Prior work has explored foundation models, supervised fine-tuning, and reinforcement learning for embodied agents, but approaches often focus on a single task level or training setting. ERA is positioned against this landscape as a framework combining embodied priors and online RL across task levels.

2 Related Work

  • Foundation-model agents progressed from prompting to code-based decision tools and supervised fine-tuning across robotic control and embodied planning.The cited works span low-level robotic control and high-level planning.Source: §2 · p. 3
  • RL is used to exploit suboptimal or failed trajectories and has been reported to foster reasoning and generalization to novel tasks.Prior research also notes that smaller LLMs and VLMs often lack basic embodied knowledge such as spatial reasoning.Source: §2 · p. 3
  • Prior studies often address only high-level or low-level tasks, motivating embodied-knowledge grounding before RL; Table 1 compares these approaches and ERA’s broader coverage.The table contrasts SFT prior types and RL features including interactive training, process-level rewards, and value learning.Source: §2 · p. 3

3 Problem Formulation

The paper models language-guided VLM agents as acting from visual observations and action history in a partially observable environment. It distinguishes high-level semantic tasks from low-level executable control commands.

3 Problem Formulation

  • A VLM agent receives visual observations and a language goal, maintains observation-action history, and selects actions through a policy.The formulation represents the task as a language-augmented POMDP with state, action, observation, transition, language, and reward components.Source: §3 · p. 3
  • The objective is to maximize expected discounted reward until the instruction is satisfied or the maximum horizon is reached.The reward is generally binary: 1 when the current state satisfies the instruction and 0 otherwise.Source: §3 · pp. 3–4
  • Low-level tasks use directly executable atomic commands, whereas high-level tasks use semantic actions decomposable into low-level primitives.The distinction aligns low-level tasks with precise perception and manipulation and high-level tasks with logical reasoning and long-term planning.Source: §3 · p. 4

4 From Priors to Policies: Training VLMs as Embodied Agents

ERA trains compact VLMs through embodied-prior supervised fine-tuning followed by online reinforcement learning. This section describes the three prior sources and an RL pipeline using structured agent responses, process-level rewards, and turn-level advantage estimation.

4.1.1 Trajectory-Augmented Priors

  • Trajectory-augmented priors add stepwise reasoning supervision to trajectories, with simulator-grounded visual descriptions used for low-level manipulation.The reasoning traces include visual descriptions, reflections, and step-level plans; simulator information is used because GPT-4o descriptions can be inaccurate for manipulation.Source: §4.1.1 · pp. 5–6

4.1.2 Environment-Anchored Priors

  • Environment-anchored priors add same-environment supervision beyond recorded trajectories, covering semantic action tasks and spatial grounding.For EB-ALFRED, datasets include masked-action modeling and action-sequence reordering; for EB-Manipulation, they include absolute, relative, and combined coordinate grounding.Source: §4.1.2 · p. 6

4.1.3 External Knowledge Priors

  • External knowledge priors transfer out-of-environment reasoning and grounding signals to complement environment-specific supervision.The paper uses OpenO1-SFT for high-level planning and SpaceThinker for low-level control.Source: §4.1.3 · p. 7

4.2 Online Reinforcement Learning

  • The online RL stage addresses costly agent training, long horizons with sparse rewards, and unstable token-level optimization through three tailored components.These are an efficient agent framework with self-summarization, process-level rewards, and turn-level advantage estimation.Source: §4.2 · p. 7

4.2.1 Agent Framework

  • The agent produces structured reasoning and executable actions from instructions, visual observations, and interaction history.Its reasoning contains a visual description, reflection on prior actions and feedback, and a multi-step plan; actions are formatted for the task environment.Source: §4.2.1 · pp. 7–8
  • Self-Summarization compresses interaction history into the latest thought, reducing context size from O(t) to O(1).The resulting context retains the latest thought, action, and environment feedback.Source: §4.2.1 · p. 8

4.2.2 Reward Design

  • Process-level rewards combine task success, rule-defined subgoal progress, and behavior shaping to provide learning signals at each turn.Behavior shaping penalizes invalid high-level actions and evaluates visual-grounding accuracy for low-level manipulation.Source: §4.2.2 · pp. 8–9

4.2.3 Turn-level Policy Optimization

  • Turn-level optimization treats a complete response as one action and shares its advantage estimate across the response tokens.The value function estimates state value per turn, aligning credit assignment with environment interactions rather than individual tokens.Source: §4.2.3 · p. 9

5 Experiments

Experiments evaluate ERA on high-level planning and low-level control, comparing prompting and training-based baselines and testing design choices across its prior-learning and RL stages. Results include strong overall performance and gains on unseen tasks.

5.1 Q1: What performance does ERA achieve compared to strong baselines?

  • 65.2% on EB-ALFRED and 48.3% on EB-Manipulation; ERA-3B exceeds GPT-4o by 8.4 and 19.4 points, respectively, and VAGEN by 12.4% and 25.4%.ERA establishes a new state-of-the-art among training-based agents on both tasks.Source: Table 3 · p. 11, §5.1 · pp. 11–12
  • ERA reaches 66% on EB-ALFRED’s unseen Spatial subset versus VAGEN’s 28%, while RL baselines fall behind ERA on unseen subsets across both benchmarks.Figure 4 illustrates error reflection on EB-ALFRED and correct star placement on EB-Manipulation.Source: Figure 4 · p. 10, §5.1 · p. 12

5.2 Q2: What role do different prior datasets play in agent performance?

  • Trajectory augmentation yields the largest individual unseen-task gains: +13.0% on EB-ALFRED and +13.6% on EB-Manipulation after Stage 2.Direct RL from the 3B base model scores zero on all tasks; EPL priors improve on raw trajectories and reduce the seen–unseen gap.Source: §5.2 · p. 12, Table 4 · p. 12
  • Combining Trajectory-Augmented and Environment-Anchored Priors yields 60% on unseen EB-ALFRED (+24% over baseline) and 43.8% on unseen EB-Manipulation (+21.9%).Environment-Anchored Priors improve seen and unseen EB-ALFRED tasks equally; External Knowledge Priors favor unseen tasks, but their overall gains are smaller than trajectory augmentation.Source: Table 4 · p. 12, §5.2 · p. 13
  • EPL and EPL+RL performance is often near-linearly correlated, with Pearson coefficients from 0.88 to 0.97 across seen and unseen sets.Higher-performing EPL models tend to maintain or amplify their advantage after RL fine-tuning.Source: §5.2 · p. 13, Figure 5 · p. 13

5.3 Q3: Is Self-Summarization Effective for Context Management?

  • The history-length comparison reports average input tokens and success rate across unseen subsets, framing context management as a performance-and-context comparison.The supplied caption does not report the outcome of that comparison.Source: Table 5 · p. 14

5.4 Q4: How do reward design and turn-level value impact RL performance?

  • Adding subgoal-based and behavior-shaping rewards to sparse success rewards raises EB-ALFRED success by 14% (46% →60%), versus +2.1% on EB-Manipulation.Using either dense signal alone yields limited gains, indicating a benefit from combining them, especially for long-horizon tasks.Source: §5.4 · p. 14
  • Turn-level GAE improves average success over token-level GAE by 12 points on EB-ALFRED (48% →60%) and 8.4 points on EB-Manipulation (35.4% →43.8%).Turn-level GAE treats the entire response as one action; bi-level GAE offers modest gains, but token-level value estimation remains unstable.Source: §5.4 · p. 14

6 Conclusion

ERA combines embodied prior learning with online RL, using multiple prior sources and RL design choices to improve embodied agents. The authors present this as a practical, scalable recipe for VLM-based agents.

6 Conclusion

  • Trajectory-augmented, environment-anchored, and external-knowledge priors each improve agent performance, with their combination yielding greater gains.The second stage emphasizes context management, dense reward shaping, and turn-level value learning.Source: §6 · pp. 14–15
  • The authors frame ERA’s combined contributions as a practical and scalable recipe for more powerful and efficient VLM-based agents in realistic settings.Source: §6 · p. 15

A Limitation

The paper identifies simulated-only evaluation as a limitation and reviews related VLM and RL work alongside its evaluation environments.

A Limitation

  • All evaluations use simulated environments, leaving ERA without validation on real-world systems.Simulation offers standardized, reproducible benchmarks while reducing time, cost, and safety risks, but limits real-world applicability.Source: A Limitation · p. 24

B Additional Related Work

  • The related-work discussion groups VLM research into vision-to-text, vision-to-action, and text-to-vision areas, and notes VLM use in reasoning and video processing.Source: B Additional Related Work · p. 24
  • Prior RL work spans preference-based alignment, reasoning, and multimodal alignment, including methods intended to improve sequence stability or reduce online interaction costs.Source: B Additional Related Work · p. 24

C Environment Details

  • The evaluation environments cover high-level household planning in EB-ALFRED and low-level robotic control in EB-Manipulation.EB-ALFRED uses AI2-THOR and discrete actions; EB-Manipulation uses CoppeliaSim and 7-DoF end-effector commands with automatic motion planning.Source: C Environment Details · p. 24

C.1 EB-ALFRED for High-level Planning Task

The paper describes two simulated environments spanning household-task planning and robotic manipulation, with distinct task interfaces and action spaces.

C.1 EB-ALFRED for High-level Planning Task

  • EB-ALFRED evaluates high-level household-task planning with PDDL-specified subgoals and eight parameterized action primitives.It covers seven task types and provides egocentric visual observations plus textual success or failure feedback.Source: §C.1 · p. 25

C.2 EB-Manipulation for Low-level Control Task

  • EB-Manipulation covers four manipulation categories using a 7-dimensional action vector for end-effector translation, rotation, and gripper state.Its automatic motion planning drives trajectories between poses, reducing the need for the agent to predict task-critical keypoints.Source: §C.2 · p. 25

D Implementation Details

The implementation details describe a staged prior-learning curriculum and an online PPO pipeline with task-specific rewards and optimization settings.

D.1.1 EPL Training Design Details

  • EPL trains environment-anchored priors first, trajectory-augmented priors next, and external knowledge before trajectory training when included.The order is designed to build grounded perception before step-level reasoning and planning skills.Source: §D.1.1 · p. 25

D.1.2 RL Reward Design Details

  • The RL reward combines task completion, intermediate subgoal progress, and domain-specific behavior shaping.Task completion is sparse, while subgoal rewards provide a dense signal for intermediate progress.Source: §D.1.2 · pp. 26–27
  • High-level subgoal completion earns a first-time +1.0 reward, while semantically invalid actions incur a −0.5 behavior penalty.Source: §D.1.2 · pp. 26–27
  • For manipulation, the agent receives a one-time subgoal reward when its end-effector approaches a target, and visual-description accuracy shapes reward by match ratio.Unparsable descriptions receive a zero match ratio; the reward is positive above 0.75 and negative below 0.25.Source: §D.1.2 · p. 27
  • Reward magnitudes are tuned to keep task completion primary while encouraging intermediate progress without creating subgoal-only local optima.Penalties and description bonuses should discourage errors without stifling exploration or dominating success and subgoal rewards.Source: §D.1.2 · p. 27

D.2.1 Embodied Prior Learning

  • The EPL implementation uses 500 × 500-pixel inputs, freezes the vision transformer, and trains with Adam, a cosine schedule, and 5% warm-up.The 3B EPL-only model uses two H200-140G nodes and takes approximately two hours for EB-Manipulation or five hours for EB-ALFRED.Source: §D.2.1 · p. 28

D.2.2 Reinforcement Learning

  • Online RL uses a custom VeRL-based framework with PPO and parallel environment rollouts for data collection.The framework uses turn-wise advantage estimation, γ = 0.99, λ = 0.99, and a 0.2 PPO clipping ratio.Source: §D.2.2 · pp. 28–29
  • PPO training uses a critic warmup, distinct rollout batch sizes for the two tasks, and turn-level minibatches.The actor and critic start from EPL weights; the vision tower stays frozen for EB-ALFRED but is fine-tuned for EB-Manipulation.Source: §D.2.2 · pp. 28–29

E Additional Experiments

The supplementary experiments examine visual-description quality and trajectory reasoning in low-level control. Rule-based descriptions and additional step-wise reasoning improve performance, particularly on unseen environments.

E Additional Experiments

  • The supplementary analysis also examines oracle visual descriptions, environment-anchored prior components, reward combinations, and token-, bi-level, and turn-level GAE.Source: E Additional Experiments · p. 29

E.1 Effect of Rule-based Visual Description for Low-level Control

  • Noisy or incorrect VLM-generated visual descriptions can misguide perception; the experiment separates rule-based description augmentation from adding step-wise reasoning derived from those descriptions.The setup isolates the effects of more precise visual grounding and structured reasoning.Source: §E.1 · pp. 29–30
  • 41.4% on seen and 26.1% on unseen environments with Traj-Aug, versus 39.6% and 22.9% with rule-based descriptions alone and 25.0% and 7.3% for raw trajectories.The gains from adding structured reasoning to rule-based descriptions are especially large on unseen environments.Source: Table 7 · p. 30, §E.1 · p. 30

E.2 Effect of Different Components in the Environment-anchored Prior Dataset

The experiments test how environment-anchored supervision components contribute to planning and control. Combining complementary components improves average success rates and subset performance in both benchmarks.

E.2 Effect of Different Components in the Environment-anchored Prior Dataset

  • 47.0 average success rate on EB-ALFRED with MAM and ASR combined, versus 41.0 with MAM alone and 44.0 with ASR alone.The combined gains also hold across Common (44.0) and Spatial (50.0) splits.Source: Table 8 · p. 30, §E.2 · pp. 30–31
  • 40.0 average success rate on EB-Manipulation with all three grounding types, versus 35.0–36.7 when any one grounding dimension is removed.On the Spatial split, using all three types reaches 27.1, compared with 18.8–20.8 without one dimension.Source: Table 9 · p. 30, §E.2 · p. 31
  • The ablation examines which environment-anchored supervision components contribute to embodied reasoning and generalization across tasks.Source: §E.2 · p. 30

E.3 Effect of Different Reward Design in RL

The reward-design analysis finds that dense composite rewards yield larger benefits on long-horizon planning than on shorter-horizon control. The authors associate dense-reward utility with task horizon length.

E.3 Effect of Different Reward Design in RL

  • 46% to 60% average success rate on EB-ALFRED with the composite reward, a 14-point gain over sparse success-only reward.On unseen subsets, the Spatial and Common Sense task scores increase by 16 and 12 points, respectively.Source: Table 10 · p. 31, §E.3 · p. 31
  • 41.7% to 43.8% overall on EB-Manipulation with dense rewards, a more modest gain than on EB-ALFRED.The authors hypothesize shorter horizons let turn-level GAE assign credit more effectively despite sparser rewards.Source: §E.3 · p. 31
  • The analysis associates dense-reward utility with task horizon and notes that reward components can target distinct capabilities and act synergistically.The authors emphasize composite rewards that combine positive guidance with negative constraints.Source: §E.3 · p. 31

E.4 Effect of Different GAEs in RL

The GAE comparison finds turn-level credit assignment strongest across long-horizon planning and low-level control. It aligns credit assignment with the agent’s environmental-action unit rather than generated-token granularity.

E.4 Effect of Different GAEs in RL

  • Token-level advantage estimation mismatches the interaction granularity because a sequence of generated tokens corresponds to one environmental action.The paper compares token-level, bi-level, and turn-level GAE; turn-level consistently performs better, while bi-level gains remain modest and rely on token-level value estimation.Source: §E.4 · pp. 31–32
  • 65.2% average success rate on EB-ALFRED with turn-level GAE, 6.4 points above token-level GAE; it also achieves 43.9% on EB-Manipulation.On EB-ALFRED unseen Common Sense and Spatial subsets, gains are 14 and 10 points; on EB-Manipulation Visual and Spatial tasks, gains are 12.5 and 14.6 points.Source: Table 11 · p. 32, §E.4 · p. 32

E.5 Error Analysis

This section examines how embodied prior learning and online reinforcement learning affect errors on unseen planning and manipulation tasks, alongside examples of embodied reasoning and action traces. It also presents trajectory-augmented manipulation demonstrations and environment-anchored priors for household action sequencing and object-coordinate grounding.

E.5 Error Analysis

  • EPL and RL both reduce errors, with EPL lowering all error types and RL especially reducing reasoning and planning errors; combined, ERA has the lowest rates across categories.The analysis covers unseen subsets of EB-ALFRED and EB-Manipulation.Source: §E.5 · p. 32
  • Reasoning and planning errors dominate high-level EB-ALFRED tasks, while perception and reasoning errors are more prevalent in low-level EB-Manipulation tasks.Source: §E.5 · p. 32
  • The analysis examines 100 unseen EB-ALFRED tasks and 98 unseen EB-Manipulation tasks, classifying errors as perception, reasoning, or planning.Perception errors misdescribe the current state; reasoning errors concern current-state or history reasoning; planning errors concern future steps.Source: §E.5 · p. 32

G.1 Raw Trajectory

  • The raw trajectory pairs a plan to pick up a star and place it in a red container with actions progressing from the star to the container and release.The plan specifies moving to the star, grasping and lifting it, then moving to the container and releasing it.Source: §G.1 · pp. 40–42

G.2 Trajectory-augmented Prior Dataset

  • Trajectory-augmented examples span EB-ALFRED and EB-Manipulation, showing replanning after a failed spoon pickup while holding a plate and a plan to place a star in a silver container.The revised plan first puts down the plate; the manipulation example lays out a staged sequence for grasping and moving the star.Source: §G.2 · pp. 42, 44, 48
  • A trajectory-augmented example pairs scene descriptions and stepwise reasoning with manipulation actions and successful environment feedback across a star-to-container task.The plan progresses from grasping and lifting the star to moving it over the container, lowering it, and releasing it.Source: §G.2 · pp. 50–52

G.3 Environment-Anchored Prior Dataset

  • Environment-anchored EB-ALFRED examples demonstrate masked-action completion and reordering household actions into a task-oriented sequence.The examples complete a soap-cleaning sequence with “put down the object in hand” and order remote relocation as finding the chair, picking up the remote, finding the lamp, and turning it on.Source: §G.3 · pp. 53–54
  • Environment-anchored EB-Manipulation examples cover absolute, relative, and combined grounding through prompts about object coordinates and spatial relations.Prompts ask for an object’s 3D coordinate, which object occupies a specified coordinate, the leftmost or rightmost object’s location, and whether a coordinate-identified object is leftmost or rightmost.Source: §G.3 · pp. 54–55

G.4 External Knowledge Prior Dataset

The appendix illustrates external-knowledge examples and details procedures for generating reasoning traces, visual descriptions, and mapped EB-ALFRED actions. These examples and rules show how varied data and environment-specific transformations are prepared for embodied-prior learning.

G.4 External Knowledge Prior Dataset

  • The external-prior examples include five-digit-number reasoning for EB-ALFRED and a spatial-distance question for EB-Manipulation.The counting example uses complementary counting; the spatial example reasons about snow-covered cars and a snowy street.Source: §G.4 · pp. 56–57

G.5 Prompt for Generating Reasoning Augmentation

  • GPT-4o is prompted to enrich raw trajectories with visual descriptions, reflections on past actions, and plans for following steps.The prompt asks for a next multi-step plan in a JSON object after a specified number of actions have been executed.Source: §G.5 · pp. 57–58

G.6 Rule-based Visual Description Generation for EB-Manipulation

  • The EB-Manipulation description algorithm classifies object colors by nearest neighbor in normalized RGB space and combines color, names, and spatial information into language.It extracts shape objects and their simulator names and colors, then uses paired coordinates when forming descriptions.Source: §G.6 · p. 58

Pseudocode for Generating Rule-based Visual Description

  • The pseudocode pairs simulator and real-object entries, prefixes names with colors, optionally sorts by x-coordinate, and assembles a left-to-right sentence.It preserves full names for “sponge” and “shape sorter,” while other names have their first token removed.Source: Pseudocode for Generating Rule-based… · p. 59

G.7 Rule-based action mapping for EB-ALFRED

  • A deterministic mapper converts high-level ALFRED actions into EB-ALFRED action sequences, handling composite, context-dependent, and stateful actions.Mapped sequences were executed in the simulator to verify the mappings.Source: §G.7 · p. 59

H Case Analysis

The case examples contrast situations where ERA corrects an EPL planning or target-selection error with cases where both agents identify the target. The examples span EB-ALFRED and EB-Manipulation.

EB-Manipulation

  • ERA identifies the azure triangular prism correctly, while EPL selects the yellow cube in the EB-Manipulation reflection example.Source: Figure 8 · p. 61
  • Both ERA and EPL identify the silver star as the target object in the EB-Manipulation example.Source: Figure 9 · p. 62

EB-ALFRED

  • ERA identifies SprayBottle 2, whereas EPL repeatedly locates the same SprayBottle in the EB-ALFRED planning example.Source: Figure 10 · p. 63
  • Both agents identify the FloorLamp as the target in the EB-ALFRED successful reflection example.Source: Figure 11 · p. 64

Related papers

Loading 2510.12693v1…