Temporal Representation Alignment: Successor Features Enable Emergent Compositionality in Robot Instruction Following

Anca Dragan; Bill Chunyuan Zheng; Kuan Fang; Sergey Levine; Vivek Myers

REVIEW 3 major objections 4 minor 2 cited by

Temporal Representation Alignment: Successor Features Enable Emergent Compositionality in Robot Instruction Following

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A temporal alignment loss that links current and future state representations lets a robot compose known subtasks into new multi-step tasks, with no explicit planner or reinforcement learning.

desk verdict Strong, consistent real-world results suggest temporal alignment as an auxiliary BC loss genuinely helps compositional generalization, but the formal proof in Section 3.5 should not be cited as established theory. read the letter →

arxiv 2502.05454 v2 pith:NK52BBK3 submitted 2025-02-08 cs.RO cs.LG

Vivek Myers , Bill Chunyuan Zheng , Anca Dragan , Kuan Fang , Sergey Levine This is my paper · ORCID

classification cs.ROcs.LG

keywords temporalrepresentationalignmentsuccessorfeaturescompositionalgeneralizationinstructionfollowinggoal-conditionedimitationcontrastivelearningofflineRLstitchingrobotmanipulation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a single auxiliary training objective—aligning the representations of a current state, a future state, and a language instruction in one embedding space—is enough to make a behavior-cloned robot policy compose known subtasks into long, never-before-seen tasks. The method, Temporal Representation Alignment (TRA), reaches an average success of about 77% across 13 real-world tabletop manipulation tasks, versus about 31% for the best baseline, and in the OGBench simulation benchmark it outperforms several offline RL baselines on 'stitch' tasks explicitly designed to test composition. If correct, this matters because it suggests that compositional generalization in sequential decision making does not require hierarchical planning, reward design, or reinforcement learning: a temporally structured representation, learned as an auxiliary loss over offline demonstrations, can carry the stitching. The paper also proves a bound on compositional imitation error in terms of the in-distribution error plus a term controlled by the horizon stretch factor $\alpha$, under a waypoint-factorization assumption.

What carries the argument

The load-bearing object is the temporal alignment loss $L_{\text{NCE}}(\phi(s_t), \psi(s_{t+k}))$, a symmetric contrastive objective that pulls the embedding of the current state toward the embedding of a state $k$ steps ahead, where $k$ is drawn from a geometric distribution with success probability $1-\gamma$. A second alignment loss ties the goal embedding $\psi(g)$ and the language embedding $\xi(\ell)$ into the same space, so the representation of an intermediate waypoint $w$ acts as a shared pivot: $\psi(w)$ sits near $\phi(w)$ and near $\psi(g)$, letting a behavioral-cloning policy be conditioned first on the waypoint and later on the goal without ever seeing the joint state-goal pair. That shared-pivot structure is what allows the policy to stitch subtasks. Theorem 1 turns this intuition into a quantitative claim: for optimal $\phi$ and $\psi$, the imitation error on a longer-horizon compositional distribution $D^*$ is at most the in-distribution error plus a term that depends only on the horizon stretch $\alpha$.

What would settle it

Compute the two sides of the inequality in Theorem 1 on an OGBench stitch environment where the waypoint factorization can be checked: train TRA with optimal features, measure the imitation error on the training distribution $D$ and on the longer-horizon compositional distribution $D^*$, and compare the gap to $\frac{\alpha-1}{2\alpha} + \frac{\alpha-2}{2\alpha}\mathbf{1}\{\alpha>2\}$; an observed gap larger than the bound would show that the waypoint-factorization premise does not hold.

Watch

Extended reading notes

Core claim

The paper's central claim is that temporal representation alignment makes compositional generalization emerge without planning, hierarchy, or reinforcement learning. TRA trains three encoders—$\phi$ for current states, $\psi$ for future states and goals, and $\xi$ for language instructions—with a symmetric contrastive loss that pairs $\phi(s_t)$ with $\psi(s_{t+k})$ and pairs $\psi(g)$ with $\xi(\ell)$, while the policy is trained by ordinary behavioral cloning conditioned on $\psi(g)$ or $\xi(\ell)$. Aligning $\phi(s_t)$ with $\psi(s_{t+k})$ gives an intermediate waypoint $w$ a shared representation as both a future state and a goal, so the policy can be reconditioned from $w$ to $g$ even though the joint state-goal pair was never seen in training. The paper formalizes this in Theorem 1, which bounds the error on a longer-horizon compositional version of the training distribution $D^*$ by the in-distribution imitation error plus a term depending on the horizon ratio $\alpha$. Empirically, TRA achieves about 77% average success on 13 real-world language-conditioned tasks versus 31% for the best baseline, and it outperforms value-based offline RL methods on several OGBench stitch tasks; the same pattern holds when tasks are specified by goal images rather than language.

Load-bearing premise

The bound's load-bearing premise is that the policy factorizes through inferred waypoints—that conditioning on the goal is equivalent to first reaching an intermediate state and reconditioning from there—which is essentially the compositional property TRA is supposed to produce.

Editorial extensions

If this is right

A behavior-cloning policy trained with TRA can follow multi-step instructions that combine subtasks never seen together in the training data, such as moving an object and then sweeping a towel.
The learned representation is used only as an auxiliary loss at training time; at inference the policy is the same goal- or language-conditioned policy, so no planner or value function is required.
The gains hold for goal-image conditioning as well as language conditioning, and TRA is competitive with or better than value-based offline RL on several simulated stitch tasks.
Because the loss is additive, it can be dropped into an existing BC pipeline with three encoders and two contrastive terms, without changing the action decoder.
The bound predicts that compositional error grows only mildly with the horizon stretch factor $\alpha$ when the waypoint factorization holds, rather than growing with the worst-case gap between distributions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper leaves implicit that the same alignment geometry should support waypoint interpolation: if $\psi(w)$ lies along the temporal path from $s$ to $g$, a goal-conditioned policy could be steered through any state whose embedding is intermediate, which would make TRA a candidate substrate for transitive inference and for planning by embedding interpolation.
A natural extension is to feed TRA's embeddings to an explicit waypoint proposer; enforcing the factorization assumption of Theorem 1 at inference time should tighten the bound in practice.
The time-isotropy condition suggests a testable limitation: TRA's advantage should shrink on datasets whose demonstrations spend most of their time near the goal, since the uniform marginal state distribution assumed by $D^*$ would be violated.
Because the paper reports failures under multimodal behavior with Gaussian policies, its successor-style representation may combine well with diffusion or mixture policies that can represent multiple modes.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The empirical core of this paper is credible and likely correct. TRA adds a temporal-contrastive alignment loss to behavior cloning and reports a large, consistent jump on compositionally out-of-distribution manipulation tasks in the real world (77% vs. 31% for the best baseline on language tasks), with an ablation against GRIF that isolates the temporal term. The OGBench stitch results are more mixed, but TRA is competitive with or better than stronger offline RL methods on several environments. That is a genuinely useful simplification: no planning, no hierarchy, no value function — just an auxiliary representation loss on top of BC.

What is new is the specific combination: using temporal alignment as an auxiliary loss over the same representations that condition the policy, rather than for value estimation or policy extraction as in VIP or CRL. The paper does not oversell this; it acknowledges that the gains degrade when BC itself fails at a task.

The soft spot is real and cannot be waved away. Theorem 1 is not established. As written, the proof in Appendix F.1 ends with an inequality whose right-hand side still contains ERR(pi;D*), making the claim tautological. Assumption 1 also assumes the exact waypoint-factorization property that TRA is supposed to induce, so the theorem does not derive compositionality from temporal alignment — it assumes a version of it. The time-isotropy condition on D* is also not generally compatible with the data-generating process in Eq. (13). I do not see a simple fix short of either substantially weakening the claim or replacing the theorem with a clearer statement about what the alignment loss does to the geometry of the representation space. The empirical results do not depend on the theorem, but the paper currently presents the theory as if it were load-bearing.

Minor but worth noting: the OGBench comparison tunes the alignment coefficient alpha per environment, and the real-world experiments use alpha=1; the paper is transparent about this, but it does mean the simulation comparison is less clean than the headline numbers suggest.

Verdict: this deserves a serious referee. The empirical contribution is strong enough that a flawed proof section should not sink the paper, but the theoretical claims need either major revision or removal. I would give it a conditional accept after the authors fix or drop the formal analysis and clarify the hyperparameter tuning in simulation. I'd also bring it to a reading group as a good case study in how strong empirical evidence and weak formal claims can coexist in one paper.

Referee Report

3 major / 4 minor

Summary. The paper proposes Temporal Representation Alignment (TRA), which adds an auxiliary InfoNCE temporal-alignment loss to behavior cloning, aligning state encodings with future-state, goal-image, and language encodings. The authors claim that this simple auxiliary objective induces compositional generalization without explicit planning, hierarchical structure, or reinforcement learning. The paper reports large empirical gains on real-world BridgeData-style manipulation tasks (e.g., 77% vs. 31% for the best baseline on language-conditioned tasks) and competitive results on several OGBench stitch tasks. The theoretical support is provided by Theorem 1 in Section 3.5, which purports to bound the compositional error ERR(π;D*) by the in-distribution error ERR(π;D) plus a constant. On reading the manuscript, the stress-test concern lands: Assumption 1 already assumes the compositional factorization that TRA is claimed to induce, and the proof in Appendix F.1 ends with a tautological bound against ERR(π;D*) rather than the stated bound against ERR(π;D).

Significance. If the empirical findings hold, they are significant: a temporally structured representation loss, used only as an auxiliary objective, may yield long-horizon compositional behavior in real-world manipulation and simulated goal-conditioned benchmarks without value functions or explicit planning. The paper ships real-robot evaluations with 5-10 trials per task, compares against several external baselines, and provides a public website with code and videos, which are strengths. The OGBench comparisons cover multiple seeds and baselines and show real gains on several stitch tasks. The main weakness is that the formal mechanism claimed for the effect is not established: the theorem is conditional on the very factorization it is supposed to explain, and the proof is incomplete. The empirical contribution is still valuable, but the theoretical section must be corrected or reframed before the paper can be accepted.

major comments (3)

[Section 3.5, Assumption 1 (Eqs. 16-17)] Assumption 1 asserts that the policy factorizes through inferred waypoints with weight P(s_t=w|s_{t+k}=g). This factorization is essentially the compositional structure that TRA is claimed to induce, and the manuscript does not show that the temporal alignment objective in Eq. (12) enforces this factorization. Consequently, Theorem 1 is conditional on the very property the paper says TRA produces, making the theoretical argument circular. Please either prove that TRA optima satisfy Assumption 1 under stated conditions, or restate the theorem as a conditional bound and move the emergence claim to the empirical section.
[Appendix F.1, Eq. (21)] The proof of Theorem 1 ends with the inequality '≤ ERR(π;D*) + (α−1)/(2α) + ...', not with '≤ ERR(π;D) + ...'. Since the left-hand side is ERR(π;D*), the final inequality is a tautological statement about the compositional error itself and does not establish the claimed control by the in-distribution error. The penultimate step, which passes from a waypoint-conditioned expectation to ERR(π;D*), is also unjustified. Please supply a complete, non-tautological proof or remove the theorem from the paper.
[Section 3.5, definition of D*] The time-isotropy condition p0(s1)=p0(s_t) for all t is not implied by the data-generation model in Eq. (13), and it is generally violated whenever the expert policy makes progress toward a goal, since the state marginal at later times is shifted toward goal states. Because Theorem 1 also states that D* is distributed according to Eq. (13), the definition of the compositional distribution is internally inconsistent. Please relax this premise or re-derive the result under explicit assumptions on the dynamics and expert policy that make time-isotropy plausible.

minor comments (4)

[Section 4.3] The text says 'Section 4 shows the success rates' when referring to the real-world results; the actual results are in Table 1, not in the section as a whole.
[Figure 7 caption] The caption contains a typo: 'generatlization' should be 'generalization'.
[Eq. (14)] The Gaussian expert density is written as e^{σ^2||π_E(s_t,i)−a_t,i||^2} without a normalizing constant or a negative exponent; please make the density precise or remove this display.
[Corollary 1.1] The proof of the language-conditioned corollary is dismissed as 'similar' to Appendix F.1, but the conditioning variable is different (ξ(ℓ)) and Assumption 1 has an extra marginalization over goals; a few lines of derivation would make the result checkable.

Circularity Check

2 steps flagged · score 6.0 of 10

Theorem 1 assumes the compositionality it claims to derive, and its proof ends in a tautological bound against ERR(pi; D*); the real-world and OGBench evaluations remain externally grounded.

self definitional [Section 3.5, Assumption 1 (Eqs. 16-17)]
"Assumption 1. The policy factorizes through inferred waypoints as: goals: π(a | s, g) = ∫ π(a | s, w) P(st = w | st+k = g) dw (16) language: π(a | s, ℓ) = ∫ π(a | s, w) P(st = w | st+k = g) P(st+k = g | ℓ) dw dg, (17) where denote by π(s, g) the MLE estimate of the action a."

The theorem is presented as showing that temporal alignment creates compositional generalization, but Assumption 1 simply asserts that the policy already factorizes through inferred waypoints with weight P(st=w|st+k=g). That factorization is precisely the compositional/stitching structure the method is supposed to induce: the waypoint w and the temporal alignment weight are the formal expression of composing subtasks. Since the theorem inherits this unproven structure rather than deriving it from the temporal alignment objective, the formal result cannot support the claim that TRA produces compositionality; it restates the compositional property as a premise.
other [Appendix F.1, Eq. (21), proof of Theorem 1]
"≤ ERR(π; D∗) + 1 H′ ED∗ [ (1−γ^H)/(1−γ) ] + (α−2)/(2α) 1{α>2} ≤ ERR(π; D∗) + (α−1)/(2α) + (α−2)/(2α) 1{α>2}. (21)"

Theorem 1 claims ERR(π;D*) ≤ ERR(π;D) + constants, but the last displayed inequality of its proof is ERR(π;D*) ≤ ERR(π;D*) + constants. The very quantity being bounded appears unchanged on both sides, so the inequality is tautological and holds regardless of what the policy learns or how the representations are aligned. The proof never reintroduces the in-distribution error ERR(π;D) at the final step, so the stated conclusion does not follow from the proof; the only way to reach ERR(π;D) would be to invoke the compositional factorization already assumed in Assumption 1. Thus the formal argument reduces to assuming and restating the target.

full rationale

The paper's formal mechanism for explaining compositionality is Section 3.5/Appendix F.1, and that derivation is circular twice over: Assumption 1 already contains the waypoint-factorized compositional policy that temporal alignment is claimed to induce, and the proof of Theorem 1 terminates in Eq. (21) with ERR(π;D*) on both sides, a tautology rather than a bound in terms of ERR(π;D). The additional time-isotropy condition p0(s1)=p0(st) is also in tension with the data-generation process in Eq. (13), though that is a consistency defect rather than a circular step. These defects mean the theoretical claim is not established by the paper's own equations. The empirical case is different: the real-world tasks are genuinely out-of-distribution compositions evaluated against external baselines (GRIF, LCBC, Octo, AWR), and the OGBench comparison uses published benchmark environments and baselines, so the empirical result retains independent content. The per-environment alpha choices in OGBench (Table 5) are ordinary hyperparameter selection rather than a circular derivation, and the self-citations are methodological rather than load-bearing. Overall, the central empirical claim is supported, but the formal derivation of emergent compositionality reduces by construction, which justifies a score of 6 rather than 0-2.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The theoretical result rests on Assumption 1, which assumes the compositional factorization that the paper claims to enable; this is the main extra postulate. The empirical claim uses no invented entities. The OGBench alpha and gamma are fitted per environment.

free parameters (3)

alignment coefficient alpha (OGBench) = 60 (medium locomotion), 100 (large locomotion), 20 (non-stitch)
Tuned per environment (Table 5); reported simulation results use the optimal alpha, so the comparison is not parameter-free.
discount factor gamma (OGBench) = 0.995 (large locomotion), 0.99 (other)
Set differently per environment class; controls temporal sampling horizon and is a chosen hyperparameter.
number of rephrased language annotations (BridgeData) = 5 additional per instruction
Dataset augmentation choice in Section B.1 that affects language grounding.

assumptions (4)

ad hoc to paper Assumption 1: policy factorizes through inferred waypoints, pi(a|s,g) = integral pi(a|s,w) P(s_t=w | s_{t+k}=g) dw, and the analogous form for language.
This is the compositional property the paper claims TRA induces; assuming it makes the theorem circular. Section 3.5, Eq. (16)-(17).
ad hoc to paper D* satisfies time-isotropy: marginal state distribution is uniform across the horizon.
A strong distributional assumption on the composed dataset (Section 3.5) that is unlikely to hold for BridgeData or OGBench, yet is used to derive the bound.
domain assumption Expert and policy are isotropic Gaussians with fixed variance.
Standard in prior work (Walke et al. 2023), used in Eq. (14) and the BC loss.
domain assumption gamma > 1 - 1/H and alpha > 1 in Theorem 1.
Conditions for the bound; not justified in the real-world setting.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Temporal Representation Alignment: Successor Features Enable Emergent Compositionality in Robot Instruction Following." pith.science (2026). https://pith.science/paper/NK52BBK3

@misc{pith2026250205454,
  author       = {Pith},
  title        = {Pith review of: Temporal Representation Alignment: Successor Features Enable Emergent Compositionality in Robot Instruction Following},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NK52BBK3}},
  note         = {Machine review of arXiv:2502.05454}
}

read the original abstract

Effective task representations should facilitate compositionality, such that after learning a variety of basic tasks, an agent can perform compound tasks consisting of multiple steps simply by composing the representations of the constituent steps together. While this is conceptually simple and appealing, it is not clear how to automatically learn representations that enable this sort of compositionality. We show that learning to associate the representations of current and future states with a temporal alignment loss can improve compositional generalization, even in the absence of any explicit subtask planning or reinforcement learning. We evaluate our approach across diverse robotic manipulation tasks as well as in simulation, showing substantial improvements for tasks specified with either language or goal images.

Figures

Figures reproduced from arXiv: 2502.05454 by the authors.

**Figure 2.** Aggregated performance on compositional gen [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** The tabletop manipulation setup used for the real [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Example rollouts of a task with TRA and LCBC. While TRA is able to successfully compose the steps to complete [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Aggregated success rate of using AWR as an addi [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Two environments from the OGBench suite ( [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: TRA performs compositional generatlization over a variety of tasks seen within BridgeData. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

**Figure 8.** Figure 8: Most of the failure cases came from the fact that a policy cannot learn depth reasoning, causing early grasping or [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]

**Figure 9.** Figure 9: Visualizing the bound (Eq. 18 from Theorem 1) on the compositional generalization error [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Grounded Semantic Re-Binding for Robust Instruction Generalization in Vision-Language-Action Models
cs.RO 2026-08 conditional novelty 7.0 of 10

Robots keep the right task in mind under paraphrased instructions but execute wrong actions because joint vision-language features shift; rerouting semantics through frozen T5 and retraining the action expert restores...
Mollified Value Learning
cs.LG 2026-02 conditional novelty 5.0 of 10

Mollified Value Learning regularizes offline goal-conditioned value estimates with a Feynman-Kac expectation version of the viscous HJB equation instead of a pointwise Eikonal constraint.

Reference graph

Works this paper leans on

16 extracted references · 6 canonical work pages · cited by 2 Pith papers

[2]

RT-H: Action Hierarchies Using Lan- guage

Belkhale, S., Ding, T., Xiao, T., Sermanet, P ., Vuong, Q., Tompson, J., Chebotar, Y., Dwibedi, D., and Sadigh, D. RT-H: Action Hierarchies Using Lan- guage. arXiv:2403.01823,

arXiv
[4]

open the drawer

C.2 Behavior Cloning We use the same architecture for LCBC and GCBC as in Myers et al. (2023); Walke et al. (2023). During the training process we use the same hyperparameters as TRA. C.3 Advantage Weighted Regression In order to train an AWR agent without separately implementing a reward critic, we follow Eysenbach et al. (2022) and use a surrogate for a...

work page 2023
[5]

Inference via Interpolation: Contrastive Representations Provably Enable Planning and In- ference

Eysenbach, B., Myers, V ., Salakhutdinov, R., and Levine, S. Inference via Interpolation: Contrastive Representations Provably Enable Planning and In- ference. arXiv:2403.04082,

arXiv
[6]

Learning Latent Dynamics for Planning From Pixels

Hafner, D., Lillicrap, T., Fischer, I., Villegas, R., Ha, D., Lee, H., and Davidson, J. Learning Latent Dynamics for Planning From Pixels. arXiv:1811.04551,

arXiv
[7]

Interac- tive Language: Talking to Robots in Real Time.IEEE Robotics and Automation Letters, (arXiv:2210.06407):1– 8,

Lynch, C., Wahid, A., Tompson, J., Ding, T., Betker, J., Baruch, R., Armstrong, T., and Florence, P . Interac- tive Language: Talking to Robots in Real Time.IEEE Robotics and Automation Letters, (arXiv:2210.06407):1– 8,

arXiv
[11]

Advantage-Weighted Regression: Simple and Scalable Off-Policy Reinforcement Learning

Peng, X.B., Kumar, A., Zhang, G., and Levine, S. Advantage-Weighted Regression: Simple and Scalable Off-Policy Reinforcement Learning. arXiv:1910.00177,

arXiv 1910
[13]

Universal Visual De- composer: Long-Horizon Manipulation Made Easy

Zhang, Z., Li, Y., Bastani, O., Gupta, A., Jayaraman, D., Ma, Y.J., and Weihs, L. Universal Visual De- composer: Long-Horizon Manipulation Made Easy. arXiv:2310.08581,

arXiv
[14]

B TRA Implementation In this section, we provide details on the implementation of temporal representation alignment (TRA) and its training process

12 Temporal Representation Alignment A Code and Website A website with code, additional visualizations, and videos is available at https://tra-paper.github.io/. B TRA Implementation In this section, we provide details on the implementation of temporal representation alignment (TRA) and its training process. B.1 Dataset Curation We use an augmented version...

work page 2023

Show all 16 references

[16]

This is to simulate the ResNet architecture and CLIP embeddings we use from real-world policy training. We define separate state and goal encoder ϕ(s) and ψ(g), and we modify LTRA as: LTRA = LBC({si, ai, s+ i }K i=1; π, ϕ, ψ) + αLNCE({si, s+ i }K i=1; ϕ, ψ) (22) The rest of th...

2023
[2017]

See, Plan, Predict: Language- Guided Cognitive Planning With Video Prediction

Attarian, M., Gupta, A., Zhou, Z., Yu, W., Gilitschen- ski, I., and Garg, A. See, Plan, Predict: Language- Guided Cognitive Planning With Video Prediction. arXiv:2210.03825,

arXiv
[2019]

Learning Transferable Visual Models From Natural Language Supervision

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P ., Clark, J., Krueger, G., and Sutskever, I. Learning Transferable Visual Models From Natural Language Supervision. International Conference on Machine Learning, arXiv:2103.00020,

arXiv
[2021]

Variational Empowerment as Representation Learn- ing for Goal-Conditioned Reinforcement Learning

Choi, J., Sharma, A., Lee, H., Levine, S., and Gu, S.S. Variational Empowerment as Representation Learn- ing for Goal-Conditioned Reinforcement Learning. International Conference on Machine Learning , pp. 1953–1963,

1953
[2022]

Planning With Goal-Conditioned Policies

Nasiriany, S., Pong, V .H., Lin, S., and Levine, S. Planning With Goal-Conditioned Policies. arXiv:1911.08453,

1911 arXiv
[2023]

LIV: Language-Image Representations and Rewards for Robotic Control

Ma, Y.J., Liang, W., Som, V ., Kumar, V ., Zhang, A., Bas- tani, O., and Jayaraman, D. LIV: Language-Image Representations and Rewards for Robotic Control. International Conference on Machine Learning, 2023a. Ma, Y.J., Sodhani, S., Jayaraman, D., Bastani, O., Ku- mar, V ., and...

2003 arXiv
[2024]

Learning Successor States and Goal-Dependent Values: A Mathematical Viewpoint

Blier, L., Tallec, C., and Ollivier, Y. Learning Successor States and Goal-Dependent Values: A Mathematical Viewpoint. arXiv:2101.07123,

arXiv
[2025]

Policy Adaptation via Language Optimization: Decomposing Tasks for Few-Shot Imitation

Myers, V ., Zheng, B.C., Mees, O., Levine, S., and Fang, K. Policy Adaptation via Language Optimization: Decomposing Tasks for Few-Shot Imitation. Confer- ence on Robot Learning, 2024a. Myers, V ., Zheng, C., Dragan, A., Levine, S., and Eysen- bach, B. Learning Temporal Distan...

arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [2]

RT-H: Action Hierarchies Using Lan- guage

Belkhale, S., Ding, T., Xiao, T., Sermanet, P ., Vuong, Q., Tompson, J., Chebotar, Y., Dwibedi, D., and Sadigh, D. RT-H: Action Hierarchies Using Lan- guage. arXiv:2403.01823,

arXiv

[2] [4]

open the drawer

C.2 Behavior Cloning We use the same architecture for LCBC and GCBC as in Myers et al. (2023); Walke et al. (2023). During the training process we use the same hyperparameters as TRA. C.3 Advantage Weighted Regression In order to train an AWR agent without separately implementing a reward critic, we follow Eysenbach et al. (2022) and use a surrogate for a...

work page 2023

[3] [5]

Inference via Interpolation: Contrastive Representations Provably Enable Planning and In- ference

Eysenbach, B., Myers, V ., Salakhutdinov, R., and Levine, S. Inference via Interpolation: Contrastive Representations Provably Enable Planning and In- ference. arXiv:2403.04082,

arXiv

[4] [6]

Learning Latent Dynamics for Planning From Pixels

Hafner, D., Lillicrap, T., Fischer, I., Villegas, R., Ha, D., Lee, H., and Davidson, J. Learning Latent Dynamics for Planning From Pixels. arXiv:1811.04551,

arXiv

[5] [7]

Interac- tive Language: Talking to Robots in Real Time.IEEE Robotics and Automation Letters, (arXiv:2210.06407):1– 8,

Lynch, C., Wahid, A., Tompson, J., Ding, T., Betker, J., Baruch, R., Armstrong, T., and Florence, P . Interac- tive Language: Talking to Robots in Real Time.IEEE Robotics and Automation Letters, (arXiv:2210.06407):1– 8,

arXiv

[6] [11]

Advantage-Weighted Regression: Simple and Scalable Off-Policy Reinforcement Learning

Peng, X.B., Kumar, A., Zhang, G., and Levine, S. Advantage-Weighted Regression: Simple and Scalable Off-Policy Reinforcement Learning. arXiv:1910.00177,

arXiv 1910

[7] [13]

Universal Visual De- composer: Long-Horizon Manipulation Made Easy

Zhang, Z., Li, Y., Bastani, O., Gupta, A., Jayaraman, D., Ma, Y.J., and Weihs, L. Universal Visual De- composer: Long-Horizon Manipulation Made Easy. arXiv:2310.08581,

arXiv

[8] [14]

B TRA Implementation In this section, we provide details on the implementation of temporal representation alignment (TRA) and its training process

12 Temporal Representation Alignment A Code and Website A website with code, additional visualizations, and videos is available at https://tra-paper.github.io/. B TRA Implementation In this section, we provide details on the implementation of temporal representation alignment (TRA) and its training process. B.1 Dataset Curation We use an augmented version...

work page 2023