TEACH: Explicit Guidance in AI & Robotics

• TEACH is a research domain where teaching is explicitly used to guide learning, from conversational agents to teach-and-repeat robotics.
• It integrates dynamic loss optimization, curriculum strategies, and human-AI dialogue to enhance learning efficiency and tailor pedagogy.
• Real-world applications include teacher–student models in ML, embodied action in AI agents, and route transfer systems in autonomous robotics.

Searching arXiv for the primary paper and closely related TEACH/teaching papers to ground the article in current literature. In contemporary AI, robotics, and educational-technology research, the label TEACH denotes several distinct but structurally related lines of work in which teaching is treated as an explicit mechanism rather than a background metaphor. The literature includes conversational teachable agents in which a human student teaches a novice agent, teacher–student optimization frameworks in machine learning, embodied dialogue benchmarks for household-task completion, teach-and-repeat navigation systems, and platforms for scalable human tutoring (Love et al., 2022, Fan et al., 2018, Krajnik et al., 2017). Across these lines, teaching can refer to the selection of instructional content, the adaptation of loss functions or goals, the orchestration of dialogue, or the replay of a previously taught trajectory.

1. Scope of the term in the literature

The term is not attached to a single architecture or task. Titles in the literature use TEACH, TEACh, TeachNow, and teach-and-repeat for research programs that share an emphasis on guided adaptation, but differ in domain, objective, and evaluation protocol (Padmakumar et al., 2021, Yang et al., 2 Aug 2025, Chaudhary et al., 28 Dec 2025).

Usage	Domain	Operational meaning
conversational teachable agent	learning-by-teaching	a student teaches a novice virtual agent
Learning to Teach / dynamic loss teaching	machine learning	a teacher model shapes data, loss, or hypothesis space
TEACH: Text Encoding as Curriculum Hints	scene text recognition	ground-truth text is injected during training and progressively masked
TEACH: Temporal Variance-Driven Curriculum	goal-conditioned RL	a teacher prioritizes goals with high temporal variance in confidence
TEACh	embodied AI	Commander–Follower dialogue for household task completion
teach-and-repeat	robotics	a robot replays a previously taught route or trajectory

This range of usage shows that “teaching” is operationalized at multiple levels: as pedagogy between human and agent, as meta-optimization over another learner, as curriculum generation, as dialogue-mediated cooperation, and as route transfer from one traversal to another. A plausible implication is that the term functions less as a single technical category than as a family resemblance across systems that explicitly model guidance.

2. Learning-by-teaching and the conversational teachable agent

A conversational teachable agent is a pedagogical agent that takes the role of a novice learner while the human student becomes the teacher. In the Curiosity Notebook platform, this framing is justified by the Protégé effect: when students believe they are teaching someone else, they tend to invest more effort, organize their knowledge better, and often learn more deeply themselves. The system extends an earlier selection-based interface with a more conversational natural-language mode, so that students can type teaching utterances rather than only select prewritten sentences. The same rocks-and-minerals classification curriculum was used in both conditions, with six rocks per condition, a short article and image for each rock, and three teaching actions—Describe, Explain, and Compare—together with the ability to quiz the agent or inspect what it had learned. The interface used mixed-initiative dialogue, so that initiative alternated between user and agent (Love et al., 2022).

The study compared Alpha, a sentence-selection condition, with Gamma, a text-input condition that encouraged paraphrasing. It recruited 46 participants; 44 were included in the learning analysis for sentence selection and 41 for text input after missing data and outlier removal. Participants completed an introductory walkthrough, a pre-test on rock classification, up to 20 minutes with each agent, a post-test after each condition, a user-experience survey, a direct comparison survey, and a post-experiment interview. The main learning measure was the difference between pre-test and post-test scores, with an additional normalization by teaching attempts because typed responses took longer. Teaching actions took about 72.5 s in sentence selection and about 127.0 s in text input. Raw score gains were slightly higher for sentence selection, but the difference was not statistically significant; when gains were normalized by teaching attempts, the advantage shifted to text input and became statistically significant.

The affective findings were more differentiated than the learning findings. Behavioral engagement, measured through quizzes, notebook checks, and navigation clicks normalized by teaching attempts, showed no significant differences. For affective engagement, participants rated sentence selection as slightly more enjoyable, text input as requiring more effort, text input as more useful for their own learning, and stress as not significantly different. After experiencing both agents, more participants said they would prefer to interact with Gamma again and judged it as more helpful for their learning. The paper’s interpretation is that typed paraphrasing increases the depth of cognitive processing while imposing a time cost.

A secondary analysis quantified paraphrasing effort with Sentence-BERT embeddings and cosine similarity between each student utterance and the relevant source sentence or sentences. Higher cosine similarity indicated less paraphrasing; lower cosine similarity indicated more paraphrasing. Greater paraphrasing effort was associated with better learning outcomes, and for rock-specific learning outcomes the reported correlation was statistically significant: $r(46) = -0.354, \ p = .0233$. No significant relationship was found between paraphrasing amount and affective engagement measures. The main stated limitation was that the study measured only short-term recall, not long-term retention.

3. Teacher–student optimization in machine learning

In machine learning, teaching is formalized as an optimization problem in which a teacher model shapes the learning conditions of a student model. In Learning to Teach, the student is defined by

$$\omega^*=\arg\min_{\omega\in \Omega}\sum_{(x,y)\in D}L(y,f_\omega(x)) \overset{\Delta}{=}\mu(D,L,\Omega),$$

and the teacher seeks to optimize the student’s eventual test performance by choosing training data $D$, loss $L$, and hypothesis space $\Omega$:

$$\min_{D,L,\Omega}\mathcal{M}(\mu(D,L,\Omega), D_{test}).$$

The teacher policy $\phi_\theta: S \rightarrow \mathcal{A}$ is trained with REINFORCE. In the paper’s main DNN experiments, the teacher selects which instances in each mini-batch are kept or discarded. Evaluations on MNIST, CIFAR-10, and IMDB with MLP, ResNet, and LSTM students reported comparable accuracy while using about 45% of the training data on MNIST, about 50% on CIFAR-10, and about 75% on IMDB relative to NoTeach. On IMDB with an LSTM student, final test accuracy was 88.54% for NoTeach, 88.80% for SPL, and 89.46% for L2T (Fan et al., 2018).

Learning to Teach with Dynamic Loss Functions moves the teacher’s intervention from data selection to objective generation. Instead of a fixed loss, the student optimizes a stage-dependent loss $l_{\Phi_t}$ whose coefficients are produced by a teacher model:

$\Phi_t = \mu_\theta(s_t).$

The student update is

$$\omega_{t+1} = \omega_t - \eta_t \frac{\partial L_{\mu_\theta(s_t)}(f_{\omega_t}, D_{train}^t)}{\partial \omega_t}.$$

The teacher is optimized through the student’s training trajectory by Reverse-Mode Differentiation, using a smooth surrogate for the target metric rather than policy optimization. In image classification, the method reduced MNIST error for an MLP from 1.94% to 1.69% and for LeNet from 0.98% to 0.77%. In IWSLT-14 German-English neural machine translation, BLEU improved from 27.28 to 29.52 for LSTM-1, from 30.86 to 31.75 for LSTM-2, and from 34.01 to 34.80 for a Transformer (Wu et al., 2018).

A related shift from how to teach to what should be taught appears in Teaching What You Should Teach (TST) for knowledge distillation. TST searches for augmented samples that the teacher can classify well but the student cannot, using a neural-network-based data augmentation module with priori bias. The training is alternating: Stage I fits meta-encoders to manual augmentation policies, Stage II optimizes magnitudes and probabilities to expose teacher-strength/student-weakness samples, and Stage III performs distillation on original and augmented data. The paper reports state-of-the-art performance on almost all teacher–student pairs across CIFAR-10, ImageNet-1k, MS-COCO, and Cityscapes, and emphasizes that freezing the augmentation priors is important for stability (Shao et al., 2022).

4. Curriculum mechanisms under the TEACH label

In scene text recognition, TEACH: Text Encoding as Curriculum Hints is a training-only strategy that injects the ground-truth text into the decoder as auxiliary input and then progressively masks it away. The decoder receives

$$\omega^*=\arg\min_{\omega\in \Omega}\sum_{(x,y)\in D}L(y,f_\omega(x)) \overset{\Delta}{=}\mu(D,L,\Omega),$$0

where $$\omega^*=\arg\min_{\omega\in \Omega}\sum_{(x,y)\in D}L(y,f_\omega(x)) \overset{\Delta}{=}\mu(D,L,\Omega),$$1 are visual tokens and $$\omega^*=\arg\min_{\omega\in \Omega}\sum_{(x,y)\in D}L(y,f_\omega(x)) \overset{\Delta}{=}\mu(D,L,\Omega),$$2 is the possibly masked text embedding sequence. The masking ratio is driven by the previous batch loss,

$$\omega^*=\arg\min_{\omega\in \Omega}\sum_{(x,y)\in D}L(y,f_\omega(x)) \overset{\Delta}{=}\mu(D,L,\Omega),$$3

so high loss implies less masking and more label guidance, while low loss implies more masking and less label guidance. The paper describes a three-stage curriculum: full label guidance, loss-aware masking, and finally fully visual recognition. TEACH is explicitly training-only, requires no external pretraining, and adds no inference overhead. Under synthetic training on MJSynth + SynthText, PARSeq + TEACH reported 97.5% on IIIT5K, 97.3% on IC13, 83.4% on IC15, 90.7% on SVTP, and 93.8% on CUTE; under real-data training with 3.3M real images, it reported 99.2% on IIIT5K, 97.5% on SVT, 98.1% on IC13, 90.8% on IC15, 96.0% on SVTP, and 97.6% on CUTE. On harder datasets, PARSeq+TEACH achieved 84.2% total accuracy in the real-data setting (Yang et al., 2 Aug 2025).

In reinforcement learning, TEACH: Temporal Variance-Driven Curriculum addresses multi-goal sparse-reward goal-conditioned RL by introducing a student–teacher paradigm in which the teacher prioritizes goals with high temporal variance in a policy-confidence score derived from the Q-function:

$$\omega^*=\arg\min_{\omega\in \Omega}\sum_{(x,y)\in D}L(y,f_\omega(x)) \overset{\Delta}{=}\mu(D,L,\Omega),$$4

Learning progress is estimated from the temporal variance of this score over a window of size $$\omega^*=\arg\min_{\omega\in \Omega}\sum_{(x,y)\in D}L(y,f_\omega(x)) \overset{\Delta}{=}\mu(D,L,\Omega),$$5, and goals are sampled proportionally to that variance. The paper also gives a theoretical connection between Q-value changes and policy evolution:

$$\omega^*=\arg\min_{\omega\in \Omega}\sum_{(x,y)\in D}L(y,f_\omega(x)) \overset{\Delta}{=}\mu(D,L,\Omega),$$6

The experimental instantiation used DDPG + HER, with evaluation on 11 tasks: 8 robotic manipulation tasks and 3 maze navigation tasks. Training used 1M timesteps for robotics tasks, 400K timesteps for maze tasks, evaluation on random goals over 20 episodes, and averaging over 7 random seeds. The reported result is consistent and notable improvement over HER-IID, VDS, SPaCE, and ProCurl, together with relative robustness to the temporal-window choice (Chaudhary et al., 28 Dec 2025).

5. Embodied, social, and human-centered teaching systems

TEACh: Task-driven Embodied Agents that Chat studies teaching as dialogue-guided embodied action. It introduced a dataset of 3,047 successful gameplay sessions collected in AI2-THOR, with over 45k utterances and 12 task types ranging from Water Plant to Prepare Breakfast. Each session involves a Commander, who has oracle tools such as Progress Check and SearchObject, and a Follower, who navigates and manipulates objects but must rely on the Commander’s language guidance. TEACh defines three benchmarks—Execution from Dialogue History (EDH), Trajectory from Dialogue (TfD), and Two-Agent Task Completion (TATC)—and baseline results show the difficulty of the setting: for TfD, the reported Episodic Transformer success rates were 1.02% on validation seen, 0.48% on validation unseen, 0.51% on test seen, and 0.17% on test unseen (Padmakumar et al., 2021).

In COACH: Cooperative Robot Teaching, the robot is a teaching partner for physical cooperative tasks. The target task is modeled as a two-player cooperative Markov game,

$$\omega^*=\arg\min_{\omega\in \Omega}\sum_{(x,y)\in D}L(y,f_\omega(x)) \overset{\Delta}{=}\mu(D,L,\Omega),$$7

and, under a moderate assumption about the student model, the teaching problem reduces to a POMDP whose hidden state includes the student policy. The teacher’s reward favors reducing the distance between the student policy and a target policy while penalizing teaching cost. For approximate online teaching, the paper decomposes the task into sub-skills and uses Item Response Theory, with mastery probability

$$\omega^*=\arg\min_{\omega\in \Omega}\sum_{(x,y)\in D}L(y,f_\omega(x)) \overset{\Delta}{=}\mu(D,L,\Omega),$$8

In Overcooked-AI and in a real-robot Cooperative Ball Maze, the paper reports that a student-aware teacher outperformed fully assistive and random teaching, and emphasizes that fully assistive behavior can make students “lazy” (Yu et al., 2023).

At MOOC scale, TeachNow addresses a different bottleneck: the coordination overhead of one-to-one help. Its workflow is teacher-initiated. A volunteer teacher clicks a button; the system searches for students who were online within the past minute, actively working on a homework problem in the IDE, and not offered help recently in the past 24 hours; one such nudgable student is randomly selected and given 30 seconds to accept or reject; and, if accepted, both enter a shared session with video, audio, chat, and a shared editor. In a six-week CS1 MOOC with around 9,000 students and 600 volunteer teachers, 102 volunteers tried TeachNow, 679 tickets were initiated, 411 were accepted, 1,056 unique students were nudged, 375 unique students were helped, and teachers delivered over 12,300 minutes of 1:1 help. The abstract reports that TeachNow sessions increased student course retention rate by almost 15%, while the text reports 40% higher course progression and that almost 30% more students in the control group dropped out one week after the session (Malik et al., 2024).

The same concern with scalable, pedagogically constrained assistance appears in DebugTA, which frames programming support as a Debugging and Teaching (DT) task. Given an erroneous program, a problem description, a pool of standard solutions, and compiler feedback, the system retrieves a structurally similar accepted solution with BM25, aligns variable names through Variable Substitution, uses an External Compiler for syntax diagnosis, and generates modification suggestions rather than a direct repair. Evaluation uses a student simulator–teacher interaction paradigm in which StuBot iteratively revises code based on the teacher’s suggestions. On CodeApex, ACMOJ, and Code4Bench, the reported result is that DebugTA consistently improved both AC Rate and AC@all over direct debugging, debugging with standard code, self-debug explanation, self-debug trace, and direct teaching, while reducing plagiarism relative to baselines that directly expose the reference solution (Fu et al., 13 Oct 2025).

A further extension of teaching beyond human–AI dyads appears in the Moltbook study of OpenClaw agents teaching one another. The platform had 2.45 million registered agents at the time of analysis, and the study examined 28,683 posts after filtering automated spam. It found an 11.4:1 statement-to-question ratio, learning-oriented content receiving about 3x more engagement than other content, and a qualitative response taxonomy in which validation accounted for 22%, knowledge extension for 18%, application for 12%, and metacognitive reflection for 7%. The authors were careful to describe these as peer-learning behaviors, while leaving open whether the agents “learn” cognitively in a human sense (Chen et al., 16 Feb 2026).

6. Teach-and-repeat as a robotics interpretation of teaching

In robotics, teach-and-repeat uses “teaching” in a procedural rather than pedagogical sense. In Navigation without localisation, the teach phase records control commands and visual cues, and the repeat phase replays learned velocities while correcting only heading with camera-based feedback. The key claim is that explicit localisation is unnecessary: if heading error is stabilized, position error does not diverge over time. The method is monocular, does not require camera calibration, and can learn and traverse arbitrarily-shaped paths in realistic indoor and outdoor conditions (Krajnik et al., 2017).

Later systems specialize this pattern for different sensing and control assumptions. Ultra-Wideband Teach and Repeat replaces vision with sparse UWB ranging, an IMU, and height sensing. During the teach pass, a robot with 3 UWB transceivers builds a locally consistent map relative to fixed anchors at unknown positions; during the repeat pass, it retraces the same trajectory. In simulation, the paper reports sub-metre tracking error, and across 1000 Monte Carlo trials the repeat tracking RMSE was consistently below 1 m, with an average of 0.26 m (Shalaby et al., 2022). Robust Visual Teach and Repeat for UGVs Using 3D Semantic Maps instead combines ORB-SLAM and YOLOv3 to build a semantic 3D map from a fixed monocular camera. It reported forward start-point and end-point distances of 2.65 ± 1.34 m and 0.41 ± 0.26 m, and backward start-point and end-point distances of 1.94 ± 1.40 m and 0.60 ± 0.20 m, while supporting both forward and backward repetition and showing robustness to moderate environmental change (Mahdavian et al., 2021).

The most elaborate version in the supplied literature is Teach-Repeat-Replan, which treats the human-taught route as a source of topological intent rather than a trajectory to be tracked literally. A human-piloted trajectory, which can be arbitrarily jerky, is converted into a topologically equivalent trajectory that is smooth, safe, and kinodynamically feasible, with local replanning used to avoid unmapped or moving obstacles during replay. The system combines global and local perception, corridor generation with convex polyhedra, Bézier-curve global planning, and B-spline local replanning, and was validated in simulation as well as challenging indoor and outdoor aggressive-flight settings (Gao et al., 2019).

Taken together, these works show that the research label TEACH spans at least three technical regimes: teaching for human learning, teaching as meta-optimization over another learner, and teaching as route specification for autonomous systems. The recurring design tension is similar across regimes. Stronger guidance often improves per-interaction efficiency—typed paraphrasing in teachable agents, dynamic loss design in supervised learning, temporally focused goal selection in RL, or topologically constrained replay in robotics—but this guidance also introduces tradeoffs in time, observability, robustness, or interpretability. The literature therefore treats teaching not as a generic synonym for supervision, but as a structured intervention whose form, timing, and granularity materially affect learning and execution.