REBot: A Multi-Domain Research Disambiguation

• REBot is a reusable label applied to unrelated systems in multi-session dialogue, robotic video synthesis, quadrupedal control, and academic regulation advising.
• Each REBot instance employs distinct methodologies, from memory compression in chatbots to reinforcement learning in dynamic obstacle evasion.
• Practical evaluations demonstrate significant gains and specialized performance metrics tailored to each domain’s unique challenges.

Searching arXiv for exact and variant uses of “REBot” to disambiguate the topic and ground the article in the relevant papers. REBot is not a single research system but a reused name that appears in several unrelated arXiv works across natural language processing, robot learning, quadrupedal control, and academic regulation advising. In current usage, the name refers at least to a multi-session open-domain chatbot called “REmember ChatBOT” (Jang et al., 2023), a real-to-sim-to-real robotic video synthesis pipeline for adapting vision-language-action models (Fang et al., 15 Mar 2025), a quadrupedal reflexive evasion framework for instantaneous dynamic obstacle avoidance (Xu et al., 8 Aug 2025), and a regulation-advising chatbot built around CatRAG (Ma et al., 2 Oct 2025). Any technical discussion of REBot therefore requires explicit disambiguation by domain and paper.

1. Naming, scope, and disambiguation

The most persistent misconception is that REBot denotes one architecture or one research lineage. The arXiv record instead shows multiple independent systems that share the same or nearly the same surface form. The conversational ReBot is an open-domain dialogue model for long-term multi-session interaction (Jang et al., 2023). The robotic-learning ReBot is a data-generation and domain-adaptation system for VLA models such as Octo and OpenVLA (Fang et al., 15 Mar 2025). The quadrupedal REBot is a control framework for low-latency obstacle evasion (Xu et al., 8 Aug 2025). The academic-advising REBot is a retrieval-and-reasoning chatbot for university regulations (Ma et al., 2 Oct 2025).

This reuse of the name is not merely cosmetic. The systems differ in task definition, inputs, outputs, training signals, and evaluation methodology. One is a text dialogue model conditioned on relationship and elapsed time; one is a synthetic data pipeline that preserves real action labels while changing foreground appearance; one is a finite-state reflex controller trained with reinforcement learning; and one is a category-routed RAG/GraphRAG system for question answering. The shared label therefore has no stable technical meaning outside its local paper context.

2. ReBot as a multi-session dialogue model

In “Conversation Chronicles: Towards Diverse Temporal and Relational Dynamics in Multi-Session Conversations” (Jang et al., 2023), ReBot is the proposed open-domain chatbot for multi-session conversation, rather than single-session dialogue. The setting is structured around episodes of 5 sessions, with long-term context defined by prior conversational content, a labeled time interval between sessions, and a predefined speaker relationship. The associated dataset, Conversation Chronicles, contains 200K episodes, 5 sessions per episode, 1M dialogue sessions, 11.7M turns, an average of 11.7 turns per session, and an average of 18.03 words per turn. Time intervals are represented by five coarse labels—“a few hours later,” “a few days later,” “a few weeks later,” “a few months later,” and “a couple of years later”—and speaker relationships are drawn from 10 types, including “Co-workers,” “Husband and Wife,” “Patient and Doctor,” and “Athlete and Coach.”

The model has two modules: a chronological summarization module and a dialogue generation module. The summarizer uses T5-base with 222M parameters, while the generator uses BART-large with 406M parameters, for about 628M, reported as about 630M parameters total. The generation problem is formalized as

$P(c \vert r, t, s, h),$

where $c$ is the next utterance, $r$ is the speaker relationship, $t$ is the time interval, $s$ is the summary of prior session content, and $h$ is the current dialogue history. The serialized input explicitly includes a relationship tag, a time marker, a prior-session summary, and within-session user–bot turns.

The central architectural claim is that long-term conversational continuity is handled through memory compression rather than full raw-history concatenation. Previous sessions are summarized into compact chronological summaries, and the generator conditions on those summaries together with relationship and elapsed time. Human evaluation of the generated dialogue episodes reports Engagingness 4.78, Humanness 4.74, Memorability 4.14, and Overall 4.55. The chronological summarization module itself receives 4.3 / 5 on human evaluation over 3K generated summaries. This suggests that, within this line of work, REBot names a comparatively lightweight long-context chatbot designed around explicit temporal and relational conditioning rather than generic conversation memory.

3. ReBot as a real-to-sim-to-real data engine for VLA adaptation

In “ReBot: Scaling Robot Learning with Real-to-Sim-to-Real Robotic Video Synthesis” (Fang et al., 15 Mar 2025), ReBot denotes a data-scaling and domain-adaptation system for vision-language-action (VLA) models. The motivating problem is the high cost of real robot data collection and the “last-mile adaptation” difficulty that remains even after large-scale pretraining. ReBot addresses this with a real-to-sim-to-real pipeline: start from a real robot trajectory, replay that trajectory in simulation while replacing the manipulated object, then composite the simulated robot/object motion back onto an inpainted real background. The formal dataset notation in the paper defines a real dataset as

$$\mathcal{D} = \{\mathbf{\tau}_{i}\}_{i=1}^M, \qquad \mathbf{\tau}_i= \{\mathbf{o}_{t}, \mathbf{a}_{t}, \mathcal{L}\}_{t=1}^T,$$

and synthesized episodes as

$$\mathbf{\tau}'_j = \{\mathbf{o}'_t, \mathbf{a}_t, \mathcal{L}'\}_{t=1}^T.$$

The action sequence $\mathbf{a}_t$ is preserved, while the video frames and instruction are changed to match the substituted object and target domain.

The pipeline has three stages. In Real $\rightarrow$ Sim, ReBot constructs aligned digital twins of the robot, cameras, and table; infers table height from the initial frame using metric depth and GroundingDINO with the text prompt "table"; replays the real action sequence in Isaac Sim 4.1 and Isaac Lab; and validates replay success by monitoring object–gripper Cartesian distance between $c$0 and $c$1. In background inpainting, it segments the robot with GroundedSAM2, segments the manipulated object using a 2D point prompt projected from the estimated 3D object position, propagates masks with SAM2, and removes robot and object with ProPainter. In Sim $c$2 Real, it composites the simulated robot and manipulated object onto the inpainted real background and edits the language instruction to match the substituted object or container.

Empirically, the system is evaluated with Octo and OpenVLA. In SimplerEnv with the WidowX robot, ReBot improves in-domain success from 16.0% to 23.2% for Octo and from 1.1% to 22.9% for OpenVLA, corresponding to gains of 7.2 percentage points and 21.8 percentage points. Out-of-domain average success rises from 6.5% to 26.4% for Octo and from 0.7% to 11.1% for OpenVLA. In real-world Franka Panda experiments, where all models are additionally finetuned with 220 real episodes, ReBot increases average success from 8% to 25% for Octo and from 25% to 45% for OpenVLA. Video quality is evaluated with VBench: ReBot reports motion smoothness 99.2%, background consistency 92.2%, average temporal quality 93.0%, compared with 96.1% for real videos, and imaging quality 66.4%, compared with 70.1% for real videos. In this paper, then, ReBot is not a robot controller but a hybrid data engine that uses simulation for object diversity while keeping real actions and real backgrounds.

4. REBot as a quadrupedal reflexive evasion framework

In “REBot: Reflexive Evasion Robot for Instantaneous Dynamic Obstacle Avoidance” (Xu et al., 8 Aug 2025), REBot is a control framework for quadrupedal robots that must evade fast-moving obstacles when conventional navigation-and-replanning methods have too little reaction time. The problem setting is instantaneous dynamic obstacle avoidance, particularly in the regime where reaction time is less than 1.5 s. The framework is organized as a finite-state machine with three stages: Normal, Avoidance, and Recovery. The system switches from Normal to Avoidance when the obstacle is approaching the robot, with the condition

$c$3

It switches from Avoidance to Recovery when instability is detected, using thresholds on body orientation, joint velocity, or base height:

$c$4

The control signal is a 12-dimensional joint target angle vector,

$c$5

produced by learned policies. Collision is defined through the signed distance between obstacle center and robot bounding box,

$c$6

The avoidance policy is trained with reinforcement learning and a reward of the form

$c$7

where the regularization terms penalize energy use and abrupt contact-force changes, and the adaptive terms encourage behavior diversity, threat-scaled response, and motion away from the obstacle. The recovery policy is trained separately to restore stable posture after aggressive evasive motions.

Training uses PPO in Isaac Gym with actor and critic MLPs of hidden size [512, 256, 128], 4096 parallel environments, 24 steps per environment per iteration, and up to 5000 iterations. In the critical 0.5–1.5 s reaction-time regime, REBot achieves ASR 0.65 and RSR 0.59, compared with 0.11/0.06 for ABS and 0.09/0.05 for RRL. In the 1.5–4.0 s regime, it reaches ASR 0.81 and RSR 0.74. Real-robot experiments on Unitree Go2 with OptiTrack report ASR = 56% and RSR = 53% under stick poking, ball throwing, and human foot kicks. In this usage, REBot refers to a low-latency reflex layer beneath normal locomotion, with explicit separation between immediate collision survival and post-evasion stabilization.

5. REBot as a regulation-advising CatRAG system

In “REBot: From RAG to CatRAG with Semantic Enrichment and Graph Routing” (Ma et al., 2 Oct 2025), REBot is an LLM-powered academic regulation advising chatbot for Can Tho University (CTU). Its core method, CatRAG, is a hybrid retrieval-and-reasoning framework that combines dense retrieval with graph-based context expansion over a hierarchical category-labeled knowledge graph. The system includes a lightweight intent classifier,

$c$8

which maps queries to regulation categories and routes graph retrieval to the corresponding category-specific subgraph. The paper defines a Regulation-Enriched Graph

$c$9

where $r$0 is the set of extracted entities and $r$1 is the set of typed relations, partitioned into subgraphs $r$2 aligned with predicted category labels.

The knowledge base is constructed from official CTU regulatory documents using Docling with OCR support, normalization, NER and POS tagging with Underthesea, a custom regulation dictionary with 46 abbreviation–full term pairs, PhoBERTv2 embeddings, chunk classification, and LLM-based relation extraction. The graph is stored in a Neo4j vector-enabled database and is described as a three-layer structure of categories, contexts/chunks, and entities. At query time, REBot embeds the query, classifies it, retrieves top chunks from the vector store and the routed subgraph, expands linked entities and relations, concatenates the evidence, and uses gpt-4o-mini or mistral-small-2506 for answer generation.

The intent-classification dataset reported in Table 1 contains 3,252 questions across six classes: Study and Training, Education, Dormitory, Discipline and Scholarships, Graduation, and Other. fastText is the best classifier, with Accuracy 95.75%, Precision 95.76%, Recall 95.75%, and F1 95.67%. The question-answering evaluation reports a best result of F1 = 98.89% for CatRAG + gpt-4o-mini at threshold 0.6, with Accuracy 98.48, Precision 99.33, and Recall 98.46. Average response time is 7.2577 s for CatRAG with gpt-4o-mini, compared with 5.1076 s for standard RAG. In this paper, REBot names a domain-specific advisory system in which the principal novelty lies in category-guided graph routing rather than generic chatbot prompting.

6. Related and frequently confused names

The exact term REBot should be distinguished from several similarly named systems that are explicitly presented as separate entities. ReCoBot is a 7-axis walking space manipulator for on-orbit satellite servicing and locomotion over iSSI-compatible structures (Scherzinger et al., 2022). ReorientBot is a vision-based manipulation system for specific-posed placement that combines onboard perception, learned waypoint scoring, and classical motion planning (Wada et al., 2022). ReinboT is “Reinforced robot GPT,” a VLA model that integrates dense return prediction and expectile regression; the paper explicitly states that it does not use the exact term REBot (Zhang et al., 12 May 2025). RePlan-Bot is a multi-level replanning agent for embodied instruction following, and its paper likewise states that it is not introduced as REBot (Gong et al., 25 May 2026). Restorebot is an autonomous platform for degraded rangeland restoration, and the paper does not state that “REBot” is an official abbreviation or alias (Such et al., 2023).

This broader naming pattern suggests that “REBot” functions more as a reusable label than as a stable term of art. In practice, precise interpretation depends entirely on the associated paper title, domain, and year. For conversational AI, REBot most often refers to the multi-session chatbot of Conversation Chronicles. In robot learning, it refers to real-to-sim-to-real video synthesis for VLA adaptation. In quadrupedal control, it refers to reflexive evasion. In academic QA, it refers to a CatRAG-based regulation adviser. The encyclopedia meaning of REBot is therefore best treated as a disambiguation across multiple independent research systems rather than as one canonical architecture.