From Trajectories to Instructions: Language-Conditioned Meta-Reinforcement Learning

Garvit Singla; Raghuram Bharadwaj Diddigi; Uma Maheswari Natarajan

REVIEW 4 major objections 6 minor 23 references

This paper claims that a language instruction can replace the gradient-based inner loop of MAML, adapting a policy in a single learned parameter offset.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 14:10 UTC pith:VAZ7YOHK

load-bearing objection LA-MAML replaces the MAML inner loop with a learned language-embedding offset; the idea is plausible and the paper is readable, but the experiments do not isolate the language-specific benefit and the statistics are thin. the 4 major comments →

arxiv 2607.18830 v1 pith:VAZ7YOHK submitted 2026-07-21 cs.LG cs.AI

From Trajectories to Instructions: Language-Conditioned Meta-Reinforcement Learning

Garvit Singla , Uma Maheswari Natarajan , Raghuram Bharadwaj Diddigi This is my paper

classification cs.LG cs.AI

keywords meta-reinforcement learningMAMLlanguage-conditioned adaptationinner loop replacementtask instructionsparameter offsetsBabyAIpolicy optimization

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper asks whether the expensive inner-loop adaptation of MAML—collecting trajectories and taking gradient steps—can be replaced by a simpler signal. It proposes to use the natural-language instruction that accompanies each task, computing a single parameter offset from the instruction's embedding. The resulting method, LA-MAML, adapts a global policy to a new task at inference time with no environment interaction. On seven BabyAI environments, it matches or improves on MAML and ANIL while cutting per-iteration wall-clock training time. The ablation shows the language component is responsible for the improvement.

Core claim

The central claim is that the adapted parameters for a task can be written as θ'_i = θ + δ f_φ(task), where f_φ is a learned encoder–adapter mapping from the task instruction, entirely replacing the inner-loop trajectory collection and gradient update of MAML. The outer loop then trains both the global parameters θ and the mapping parameters φ jointly so that the single-step offset produces a policy that works on the sampled task. Because the inner loop is gone, inference on an unseen task only requires embedding its instruction and adding the offset; no environment rollout is needed for adaptation.

What carries the argument

The load-bearing identity is θ'_i = θ + δ f_φ(task) (Eq. 5). Here δ is a fixed step-size parameter, θ is the shared global policy initialization, and f_φ is composed of a frozen pretrained sentence-transformer encoder plus a learned adapter network that projects the instruction embedding into the policy parameter space. This object replaces the inner-loop gradient step, and the outer loop (optimized with TRPO) ties θ and φ together so that the offset is meaningful.

Load-bearing premise

A frozen pretrained sentence transformer's embedding of a templated instruction carries enough information about the task to let a learned adapter predict useful policy-parameter offsets for held-out tasks.

What would settle it

Hold out a task whose instruction differs only in an attribute the embedding pools together (e.g., 'go to the red ball' vs 'go to the blue ball' if color is conflated); if the adapter cannot separate them, performance drops to using θ alone. Alternatively, feed a paraphrased but semantically identical instruction and measure whether the offset stays stable.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Inference-time adaptation to a new task costs one forward pass through the language encoder and adapter; no environment interaction is required.
The same global parameters can serve many tasks as long as a language embedding can separate them.
Per-iteration wall-clock training time drops because trajectory collection for inner-loop adaptation is removed.
The performance gap over a language-conditioned policy without meta-learning indicates the outer loop still contributes generalization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The single-offset idea could replace inner loops in other meta-learning settings, such as supervised few-shot learning, whenever task descriptions are available.
If the offset mapping is near-linear in the embedding, the method may be interpreted as a learned 'task vector' in parameter space, connecting to weight interpolation and task arithmetic.
The approach is likely sensitive to the quality of the frozen embedding; paraphrased or noisy instructions may break the transfer, a regime the paper explicitly leaves untested.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

LA-MAML replaces the MAML inner loop with a learned language-embedding offset; the idea is plausible and the paper is readable, but the experiments do not isolate the language-specific benefit and the statistics are thin.

read the letter

First, the takeaway: LA-MAML is a simple, sensible idea — replace MAML's inner-loop gradient adaptation with a single offset computed from a frozen sentence encoder plus a learned adapter, then train the whole thing end-to-end. The BabyAI results are directionally supportive: LA-MAML usually beats MAML and ANIL in average steps and often trains faster. The paper is worth a careful read. What is actually new: prior language-conditioned meta-RL work used language as an auxiliary reward signal or as context for gradient-based adaptation. Here the instruction embedding directly produces the full parameter offset and the gradient inner loop is gone entirely. That is a clean, plausible design, and the paper's observation that the outer loop does the heavy lifting so the inner loop can be non-gradient is fair. The ablation using theta alone at inference is also a decent sanity check: the language offset is carrying task-specific information, not just acting as a random perturbation. Soft spots, in proportion: the central comparison is confounded. LA-MAML receives a privileged signal — the instruction — while MAML and ANIL get no task descriptor and must infer everything from reward. Any advantage could come from extra information, not from the offset mechanism. The one language-conditioned baseline concatenates the instruction to the state but does not use the offset architecture or the meta-objective, so it does not control for architecture. To support the claim that this is a good meta-learning mechanism, the authors need a gradient-based inner-loop method that also sees the instruction, or a non-meta policy using the identical offset architecture. As it stands, the stress-test note is right to flag this. Also, no seeds and no significance tests: several standard deviations are large and overlapping (GoToLocal: 43.59±22.35 vs 44.19±17.27), and hyperparameters are only promised via an unverified GitHub link. These are addressable, not fatal. The paper's own limitation section concedes that natural, noisy language is untested, which is honest. The method may well transfer to messier instructions — or not — but that is exactly what an empirical paper needs to check. Bottom line: the idea is genuine, the experiments are directionally supportive, but the specific claim that language-driven offsets are better than gradient-based adaptation for the same information is not yet established. A serious referee could push the authors to add the missing controls and report statistics; that would make it a much stronger paper. I would send it to review, and I would read the revision.

Referee Report

4 major / 6 minor

Summary. The paper proposes LA-MAML, a meta-reinforcement learning method that replaces MAML's gradient-based inner-loop adaptation with a single language-conditioned parameter offset computed as theta'_i = theta + delta * f_phi(task) (Eq. 5). A frozen Sentence Transformer (all-MiniLM-L6-v2) embeds the task instruction, and a learned adapter network projects this embedding into policy-parameter offsets. The outer loop trains theta and phi jointly with TRPO (Eq. 7). The method is evaluated on seven BabyAI environments against MAML, ANIL, and a non-meta language-conditioned policy. The paper reports lower per-iteration wall-clock training time and competitive/improved evaluation performance, and an ablation shows that removing the language offset degrades performance. The source code is made available.

Significance. If validated, the contribution is useful: LA-MAML provides a concrete mechanism for making MAML-style adaptation interaction-free at deployment, and the algorithmic idea is clearly presented with pseudo-code and held-out task evaluation. The ablation comparing LA-MAML to inference with theta only is a strength, and the claims are falsifiable. However, the current experimental evidence does not yet isolate the offset mechanism from the extra information provided by language, and the statistical support is thin. The paper is honest about its limitation to templated BabyAI instructions in Section 5.4, but the abstract and conclusion over-generalize the result. With additional controls and multi-seed statistics, the contribution could be solid; as presented, the central empirical claim is not fully established.

major comments (4)

[§5.1, Table 2, §5.2] The main claim—that language-conditioned offsets can replace gradient-based inner-loop adaptation—is not isolated by the current baselines. MAML and ANIL receive no task descriptor and must infer the task from rewards, whereas LA-MAML observes the instruction; any advantage could come from the extra information rather than the offset mechanism. The only language-conditioned baseline is a policy conditioned on (s, ℓ) that is trained without the meta-objective and does not use the offset architecture. To support the claim, the authors should compare against a gradient-based inner-loop method that also receives the language instruction (e.g., MAML/ANIL with the instruction embedding concatenated to the state or as an additional input), or against a non-meta policy using the same θ + δ f_φ(T) architecture. Without such a control, §5.2 point 2 ('meta-learning drives the improvement') is not e
[Table 2, Fig. 2] The evaluation reports only means and standard deviations, with no number of seeds, no per-configuration breakdown, and no significance tests. Several headline improvements are within overlap of the reported standard deviations: in GoToLocal, LA-MAML (43.59±22.35) vs MAML (44.19±17.27) differ by 0.6 steps with SDs over 17–22. In OpenDoorLoc, LA-MAML has SD 91.36 and the reported gap may be driven by variance. The paper should report the number of seeds, confidence intervals or paired tests, and the per-configuration results. As is, the 'competitive or improved performance' claim is not statistically supported.
[§5.1, Algorithm 1] The architecture and training details needed to assess fairness and reproducibility are incomplete. The authors state 'identical network architectures and hyperparameters across all comparison methods,' but the language-conditioned policy must have a different input interface (state and instruction) than MAML/ANIL. The adapter network structure, the hidden dimension of the frozen embedding, and the value of the step-size δ in Eq. (5) are not reported. The paper should provide a full architecture table, the δ value, the number of seeds, and the key hyperparameters (TRPO KL constraint, GAE λ, K, learning rates). The GitHub link is useful but should not replace details in the paper.
[Eq. (5), Algorithm 1 lines 10–11, §4] The phrase 'replacing the inner loop trajectory collection and gradient-based updates' is stronger than what the method does during training. In Algorithm 1, K trajectories are still collected (lines 10–11) after computing θ'_i, and the outer-loop loss is computed on them. What is replaced is the use of trajectories to compute a gradient-based adaptation step, not the collection of trajectories in training. At inference, the method indeed requires no trajectories. This distinction should be stated precisely; otherwise the wall-clock comparison in Table 1 is confounded by the fact that the number of environment interactions per meta-iteration may differ across methods. Please clarify the exact number of interaction steps used by each method per meta-iteration.

minor comments (6)

[Notation] δ denotes both the step size in Eq. (5) and the temporal-difference residual in Algorithm 1 line 13. Use different symbols (e.g., η for step size).
[Algorithm 1 lines 7–9] The notation is inconsistent: the text introduces e_i, the pseudocode uses \tilde{f}_i, and then f_φ(T_i) is used for the reshaped offset. State clearly that f_φ = adapter ∘ encoder and keep one name throughout.
[Fig. 2] The figure caption does not explain the shading beyond 'standard deviation,' nor how many rollouts per point are used. Please specify the smoothing, the number of meta-iterations shown, and whether the y-axis 'average steps' is capped at the environment maximum.
[§2.1 and References] The sentence mentioning ANIL, CAVIA, and WarpGrad does not place citations next to each method consistently; [16] appears unused in this sentence. Check all citations for accuracy.
[§5.1] Environment configuration names such as 'GoToLocalS5N2' run together; please add spacing for readability.
[§5.4] The limitation about imperfect or noisy language is acknowledged, but the abstract and conclusion claim that language instructions are an 'effective and efficient substitute' without this caveat. Please temper the wording accordingly.

Circularity Check

0 steps flagged

No significant circularity: LA-MAML's parameter-offset adaptation is an empirical meta-learning objective, trained on training tasks and evaluated on held-out tasks, with no load-bearing self-citation or definitional reduction.

full rationale

The paper's central mechanism, θ′_i = θ + δ f_φ(task) (Eq. 5), is a definition of the adaptation procedure, not a prediction derived from fitted data. The meta-objective (Eq. 6) is exactly the loss minimized over training tasks; evaluation is performed on held-out tasks across configurations, which is a genuine generalization test. The language encoder is a frozen pretrained Sentence Transformer, and only the adapter and global policy are trained; nothing is fitted to the test set or renamed as a prediction. The ablation in Table 3 compares the full model to inference with θ only, demonstrating that the language component contributes empirically rather than being circular. The paper contains no self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The limitations section (5.4) concedes that noisy/imperfect language is untested, which is an acknowledged scope limitation, not circularity. Baseline fairness concerns—e.g., LA-MAML receiving privileged language input while MAML/ANIL do not—are experimental comparison issues, not circular derivation; they affect the strength of the empirical claim, but do not make any result equivalent to its inputs by construction. Therefore no circular step meets the required evidentiary standard.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The method introduces no new physical or mathematical entities. It relies on a learned adapter network and frozen language encoder, plus several unspecified hyperparameters. The main unstated assumptions are the sufficiency of the frozen sentence embeddings and the comparability of baselines.

free parameters (4)

Step-size δ (Eq. 5)
Scalar controlling how much the language offset changes the global parameters; value is not reported in the manuscript and must be tuned by hand.
GAE λ and discount γ
Required by Algorithm 1 lines 13–14; values are not listed in the experimental setup.
Number of episodes K per task
Algorithm 1 line 11 requires K; it controls trajectory collection cost and variance, but K is not reported.
TRPO KL constraint and natural-gradient settings
Outer-loop update in Eq. (7) depends on trust-region hyperparameters, which are not specified.

axioms (6)

domain assumption Each sampled task Ti comes with a natural-language instruction whose embedding is a sufficient signal for optimal task-specific policy offsets.
This is the premise of Eq. (5); the paper's Section 5.4 narrows it to templated, structured instructions.
domain assumption The frozen Sentence Transformer (all-MiniLM-L6-v2) embeds BabyAI instructions into a space where a learned adapter can produce useful policy offsets.
No analysis verifies the encoder's separation of task semantics; the adapter must compensate for any embedding insufficiency.
standard math Outer-loop natural-gradient TRPO update correctly optimizes joint parameters [θ, φ] for the meta-objective in Eq. (6).
TRPO is standard background; accepted without proof.
standard math The trajectory-based loss L_Ti computed from the adapted policy is a valid estimate of the meta-objective.
Relies on standard policy gradient and GAE assumptions.
domain assumption Identical network architectures and hyperparameters across baselines ensure a fair comparison.
Stated in Section 5.1 but not verifiable from the manuscript; if violated, performance gaps could come from capacity rather than method.
domain assumption Task distribution p(T) covers both training and held-out tasks with the same language distribution.
The generalization claim depends on held-out tasks resembling training tasks in language structure and environment dynamics.

pith-pipeline@v1.3.0-alltime-deepseek · 7892 in / 10555 out tokens · 96241 ms · 2026-08-01T14:10:06.620039+00:00 · methodology

0 comments

read the original abstract

Model-Agnostic Meta-Learning (MAML) is a widely used framework for reinforcement learning (RL) that enables efficient transfer by learning global policy parameters that can be rapidly adapted to new tasks. MAML training proceeds in two loops: an inner loop where the global parameters are adapted to task-specific parameters, and an outer loop where these task-specific parameters are evaluated and losses are back-propagated to improve the global parameters. Traditionally, the inner loop adaptation is performed by collecting trajectories from the task environment and applying gradient updates on the empirical expected return, which can be a costly operation. We note that it is the outer loop that drives the actual learning of global parameters, and therefore the inner loop adaptation mechanism need not be restricted to be gradient-based. This observation leads us to ask: Can we replace the inner loop trajectory collection and gradient update with a simpler, task-specific signal? In many practical settings, tasks are naturally accompanied by language instructions. Leveraging these instructions as a direct task-specific signal, we propose LA-MAML (Language Adapted MAML), which modifies the inner loop by adapting the global policy parameters in a single step through a learned embedding of the task instruction, replacing the inner loop trajectory collection and gradient-based updates. Experiments on the BabyAI benchmark demonstrate that LA-MAML achieves competitive or improved performance compared to baselines at a significantly lower per-iteration wall-clock training time. These results demonstrate that language instructions are an effective and efficient substitute for trajectory-based inner loop adaptation in meta RL.

Figures

Figures reproduced from arXiv: 2607.18830 by Garvit Singla, Raghuram Bharadwaj Diddigi, Uma Maheswari Natarajan.

**Figure 1.** Figure 1: Our Proposed LA-MAML trains the global parameters θ ∗ and trained encoder ϕ ∗ in an end-to-end fashion. During inference, the task adaption for task τ is obtained by simply adding θ ∗ and ϕ ∗ (τ ) The key contributions of this work are summarized as follows: 1. We propose LA-MAML, a novel meta-learning framework for RL that leverages natural language task descriptions through a learned encoder–adapter mec… view at source ↗

**Figure 2.** Figure 2: Training curves shows the convergence over meta-iterations for LA-MAML and baselines across seven environments. The shaded regions correspond to the standard deviation, while the solid lines represent the mean average steps per meta-iteration [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 8 linked inside Pith

[1]

arXiv preprint arXiv:2209.04924 (2022)

Bing, Z., Koch, A., Yao, X., Huang, K., Knoll, A.: Meta-reinforcement learning via language instructions. arXiv preprint arXiv:2209.04924 (2022)

Pith/arXiv arXiv 2022
[2]

arXiv preprint arXiv:1604.07316 (2016)

Bojarski, Mariusz and Del Testa, Davide and Dworakowski, Daniel and Firner, Bernhard and Flepp, Beat and Goyal, Prasoon and Jackel, Lawrence D and Mon- fort, Mathew and Muller, Urs and Zhang, Jiakai and others: End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316 (2016)

Pith/arXiv arXiv 2016
[3]

arXiv preprint arXiv:1810.08272 (2018)

Chevalier-Boisvert, M., Bahdanau, D., Lahlou, S., Willems, L., Saharia, C., Nguyen, T.H., Bengio, Y.: Babyai: A platform to study the sample efficiency of grounded language learning. arXiv preprint arXiv:1810.08272 (2018)

Pith/arXiv arXiv 2018
[4]

In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing

Deb, B., Hassan, A., Zheng, G.: Boosting natural language generation from in- structions with meta-learning. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 6792–6808 (2022)

2022
[5]

In: International conference on machine learning

Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: International conference on machine learning. pp. 1126–1135. PMLR (2017)

2017
[6]

arXiv preprint arXiv:1909.00025 (2019) 12 Authors Suppressed Due to Excessive Length

Flennerhag, S., Rusu, A.A., Pascanu, R., Visin, F., Yin, H., Hadsell, R.: Meta- learning with warped gradient descent. arXiv preprint arXiv:1909.00025 (2019) 12 Authors Suppressed Due to Excessive Length

Pith/arXiv arXiv 1909
[7]

The Inter- national Journal of Robotics Research40(4-5), 698–721 (2021)

Ibarz, J., Tan, J., Finn, C., Kalakrishnan, M., Pastor, P., Levine, S.: How to train your robot with deep reinforcement learning: lessons we have learned. The Inter- national Journal of Robotics Research40(4-5), 698–721 (2021)

2021
[8]

Journal of Machine Learning Research17(39), 1–40 (2016)

Levine,S.,Finn,C.,Darrell,T.,Abbeel,P.:End-to-endtrainingofdeepvisuomotor policies. Journal of Machine Learning Research17(39), 1–40 (2016)

2016
[9]

Statistics in medicine41(20), 4034–4056 (2022)

Liu, M., Shen, X., Pan, W.: Deep reinforcement learning for personalized treatment recommendation. Statistics in medicine41(20), 4034–4056 (2022)

2022
[10]

In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) (2022)

Min, S., Lewis, M., Zettlemoyer, L., Hajishirzi, H.: Metaicl: Learning to learn in context. In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) (2022)

2022
[11]

arXiv preprint arXiv:1312.5602 (2013)

Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., Riedmiller, M.: Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013)

Pith/arXiv arXiv 2013
[12]

Nature518, 529–533 (2015)

Mnih, Volodymyr and Kavukcuoglu, Koray and Silver, David and Rusu, Andrei A and Veness, Joel and Bellemare, Marc G and Graves, Alex and Riedmiller, Martin and Fidjeland, Andreas K and Ostrovski, Georg and others: Human-level control through deep reinforcement learning. Nature518, 529–533 (2015)

2015
[13]

In: Advances in Neural Information Processing Systems (NeurIPS) (2018)

Nichol, A., Achiam, J., Schulman, J.: On first-order meta-learning algorithms. In: Advances in Neural Information Processing Systems (NeurIPS) (2018)

2018
[14]

arXiv preprint arXiv:1909.09157 (2019)

Raghu, A., Raghu, M., Bengio, S., Vinyals, O.: Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157 (2019)

Pith/arXiv arXiv 1909
[15]

In: Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP)

Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert- networks. In: Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). pp. 3982–3992 (2019)

2019
[16]

arXiv preprint arXiv:1807.05960 (2018)

Rusu, A.A., Rao, D., Sygnowski, J., Vinyals, O., Pascanu, R., Osindero, S., Hadsell, R.: Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960 (2018)

Pith/arXiv arXiv 2018
[17]

In: International conference on machine learning

Schulman, J., Levine, S., Abbeel, P., Jordan, M., Moritz, P.: Trust region policy optimization. In: International conference on machine learning. pp. 1889–1897. PMLR (2015)

2015
[18]

arXiv preprint arXiv:1506.02438 (2015)

Schulman, J., Moritz, P., Levine, S., Jordan, M., Abbeel, P.: High-dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438 (2015)

Pith/arXiv arXiv 2015
[19]

Nature529, 484–489 (2016)

Silver, David and Huang, Aja and Maddison, Chris J and Guez, Arthur and Sifre, Laurent and Van Den Driessche, George and Schrittwieser, Julian and Antonoglou, Ioannis and Panneershelvam, Veda and Lanctot, Marc and others: Mastering the game of go with deep neural networks and tree search. Nature529, 484–489 (2016)

2016
[20]

In: Pro- ceedings of the 2020 Conference on Robot Learning (CoRL) (2020)

Stepputtis, S., Campbell, J., Phielipp, M., Lee, S., Baral, C., Ben Amor, H.: Language-conditioned imitation learning for robot manipulation tasks. In: Pro- ceedings of the 2020 Conference on Robot Learning (CoRL) (2020)

2020
[21]

MIT Press, 2 edn

Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT Press, 2 edn. (2018)

2018
[22]

In: Proceedings of the 2022 IEEE International Conference on Robotics and Automation (ICRA) (2022)

Yao, X., Bing, Z., Zhuang, G., Chen, K., Zhou, H., Huang, K., Knoll, A.: Learn- ing from symmetry: Meta-reinforcement learning with symmetrical behaviors and language instructions. In: Proceedings of the 2022 IEEE International Conference on Robotics and Automation (ICRA) (2022)

2022
[23]

In: Proceedings of the 36th International Conference on Machine Learning (ICML) (2019)

Zintgraf, L., Shiarlis, K., Kurin, V., Hofmann, K., Whiteson, S.: Fast context adap- tation via meta-learning. In: Proceedings of the 36th International Conference on Machine Learning (ICML) (2019)

2019

[1] [1]

arXiv preprint arXiv:2209.04924 (2022)

Bing, Z., Koch, A., Yao, X., Huang, K., Knoll, A.: Meta-reinforcement learning via language instructions. arXiv preprint arXiv:2209.04924 (2022)

Pith/arXiv arXiv 2022

[2] [2]

arXiv preprint arXiv:1604.07316 (2016)

Bojarski, Mariusz and Del Testa, Davide and Dworakowski, Daniel and Firner, Bernhard and Flepp, Beat and Goyal, Prasoon and Jackel, Lawrence D and Mon- fort, Mathew and Muller, Urs and Zhang, Jiakai and others: End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316 (2016)

Pith/arXiv arXiv 2016

[3] [3]

arXiv preprint arXiv:1810.08272 (2018)

Chevalier-Boisvert, M., Bahdanau, D., Lahlou, S., Willems, L., Saharia, C., Nguyen, T.H., Bengio, Y.: Babyai: A platform to study the sample efficiency of grounded language learning. arXiv preprint arXiv:1810.08272 (2018)

Pith/arXiv arXiv 2018

[4] [4]

In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing

Deb, B., Hassan, A., Zheng, G.: Boosting natural language generation from in- structions with meta-learning. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 6792–6808 (2022)

2022

[5] [5]

In: International conference on machine learning

Finn, C., Abbeel, P., Levine, S.: Model-agnostic meta-learning for fast adaptation of deep networks. In: International conference on machine learning. pp. 1126–1135. PMLR (2017)

2017

[6] [6]

arXiv preprint arXiv:1909.00025 (2019) 12 Authors Suppressed Due to Excessive Length

Flennerhag, S., Rusu, A.A., Pascanu, R., Visin, F., Yin, H., Hadsell, R.: Meta- learning with warped gradient descent. arXiv preprint arXiv:1909.00025 (2019) 12 Authors Suppressed Due to Excessive Length

Pith/arXiv arXiv 1909

[7] [7]

The Inter- national Journal of Robotics Research40(4-5), 698–721 (2021)

Ibarz, J., Tan, J., Finn, C., Kalakrishnan, M., Pastor, P., Levine, S.: How to train your robot with deep reinforcement learning: lessons we have learned. The Inter- national Journal of Robotics Research40(4-5), 698–721 (2021)

2021

[8] [8]

Journal of Machine Learning Research17(39), 1–40 (2016)

Levine,S.,Finn,C.,Darrell,T.,Abbeel,P.:End-to-endtrainingofdeepvisuomotor policies. Journal of Machine Learning Research17(39), 1–40 (2016)

2016

[9] [9]

Statistics in medicine41(20), 4034–4056 (2022)

Liu, M., Shen, X., Pan, W.: Deep reinforcement learning for personalized treatment recommendation. Statistics in medicine41(20), 4034–4056 (2022)

2022

[10] [10]

In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) (2022)

Min, S., Lewis, M., Zettlemoyer, L., Hajishirzi, H.: Metaicl: Learning to learn in context. In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) (2022)

2022

[11] [11]

arXiv preprint arXiv:1312.5602 (2013)

Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., Riedmiller, M.: Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013)

Pith/arXiv arXiv 2013

[12] [12]

Nature518, 529–533 (2015)

Mnih, Volodymyr and Kavukcuoglu, Koray and Silver, David and Rusu, Andrei A and Veness, Joel and Bellemare, Marc G and Graves, Alex and Riedmiller, Martin and Fidjeland, Andreas K and Ostrovski, Georg and others: Human-level control through deep reinforcement learning. Nature518, 529–533 (2015)

2015

[13] [13]

In: Advances in Neural Information Processing Systems (NeurIPS) (2018)

Nichol, A., Achiam, J., Schulman, J.: On first-order meta-learning algorithms. In: Advances in Neural Information Processing Systems (NeurIPS) (2018)

2018

[14] [14]

arXiv preprint arXiv:1909.09157 (2019)

Raghu, A., Raghu, M., Bengio, S., Vinyals, O.: Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157 (2019)

Pith/arXiv arXiv 1909

[15] [15]

In: Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP)

Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert- networks. In: Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). pp. 3982–3992 (2019)

2019

[16] [16]

arXiv preprint arXiv:1807.05960 (2018)

Rusu, A.A., Rao, D., Sygnowski, J., Vinyals, O., Pascanu, R., Osindero, S., Hadsell, R.: Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960 (2018)

Pith/arXiv arXiv 2018

[17] [17]

In: International conference on machine learning

Schulman, J., Levine, S., Abbeel, P., Jordan, M., Moritz, P.: Trust region policy optimization. In: International conference on machine learning. pp. 1889–1897. PMLR (2015)

2015

[18] [18]

arXiv preprint arXiv:1506.02438 (2015)

Schulman, J., Moritz, P., Levine, S., Jordan, M., Abbeel, P.: High-dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438 (2015)

Pith/arXiv arXiv 2015

[19] [19]

Nature529, 484–489 (2016)

Silver, David and Huang, Aja and Maddison, Chris J and Guez, Arthur and Sifre, Laurent and Van Den Driessche, George and Schrittwieser, Julian and Antonoglou, Ioannis and Panneershelvam, Veda and Lanctot, Marc and others: Mastering the game of go with deep neural networks and tree search. Nature529, 484–489 (2016)

2016

[20] [20]

In: Pro- ceedings of the 2020 Conference on Robot Learning (CoRL) (2020)

Stepputtis, S., Campbell, J., Phielipp, M., Lee, S., Baral, C., Ben Amor, H.: Language-conditioned imitation learning for robot manipulation tasks. In: Pro- ceedings of the 2020 Conference on Robot Learning (CoRL) (2020)

2020

[21] [21]

MIT Press, 2 edn

Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT Press, 2 edn. (2018)

2018

[22] [22]

In: Proceedings of the 2022 IEEE International Conference on Robotics and Automation (ICRA) (2022)

Yao, X., Bing, Z., Zhuang, G., Chen, K., Zhou, H., Huang, K., Knoll, A.: Learn- ing from symmetry: Meta-reinforcement learning with symmetrical behaviors and language instructions. In: Proceedings of the 2022 IEEE International Conference on Robotics and Automation (ICRA) (2022)

2022

[23] [23]

In: Proceedings of the 36th International Conference on Machine Learning (ICML) (2019)

Zintgraf, L., Shiarlis, K., Kurin, V., Hofmann, K., Whiteson, S.: Fast context adap- tation via meta-learning. In: Proceedings of the 36th International Conference on Machine Learning (ICML) (2019)

2019