PAPER REPORTENAll readings ↗

CLAM: Continuous Latent Action Models for Robot Learning from Unlabeled Demonstrations

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Anthony Liang; Pavel Czempin; Matthew M. Hong; Yutai Zhou; Jingzhen Wang; Erdem Bıyık; Stephen Tu

Affiliations: Thomas Lord Department of Computer Science, University of Southern California; Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California

Source: IEEE/RSJ International Conference on Intelligent Robots and Systems 2026 (arXiv journal reference) · 2505.04999 ↗ · Catalog record

Reading: 361 / 558 · 6 original figures & tables · ~19 min ·

1. Paper overview

In one sentence: CLAM makes observation-only robot imitation executable by jointly grounding continuous latent actions, but still depends on labeled transition coverage and task-specific robot demonstrations. e01e02e03e04e05e11e12

At a glanceWhat to know
Research problem
Source description

How can a robot imitate task-specific demonstrations whose actions are withheld? CLAM separates a large observation corpus, a small action-labeled grounding set, and observation-only expert demonstrations. Experiments isolate action inference using robot data in a single-task, single-embodiment setting. Section III allows some task-specific demonstrations within the labeled mixture; the strongest “no expert labels” wording must therefore be read alongside the purely suboptimal-data ablation. e02e14

Core mechanism
Source description

Combines continuous latent inverse dynamics with a jointly trained action decoder that remains in the execution pipeline. e03e04e05

A key reported resultMetaWorld image-based manipulation: ST-ViViT-CLAM: 76%, reported in the results text.

Average task success rate. Four manipulation tasks; Figure 2 specifies 50 evaluation rollouts across three seeds and 100 labeled trajectories.

LAPO: 9%; LAPA: 20%, under the paper’s limited, non-fully-expert labeled-data regime. This is executed simulated control. The caption and appendix disagree on labeled-data count; the caption also names variants absent from the plotted legend. e08e10e17

Reading caution
Reader analysis

Internet-video robustness and human-to-robot transfer remain future work. Expert robot demonstrations are still needed, with labels withheld. The conclusion’s broad “without expert teleoperation” wording exceeds this experimental scope. e02e15

Core contributions

  • Source description

    Combines continuous latent inverse dynamics with a jointly trained action decoder that remains in the execution pipeline. e03e04e05

  • Author claim

    The authors argue that this combination makes latent-action relabeling practical for continuous control; the factorial MetaWorld ablation directly tests these two choices. e11

Figure 1. Reconstruction learns transition codes; joint action supervision makes those codes executable. Original paper, p. 2 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Follow Stage 1 from the observation pair into the green latent IDM. Its output z_t branches toward the blue FDM and the purple action decoder. The FDM reconstructs the next observation, while the decoder predicts the motor command on labeled transitions. Algorithm 1 clarifies the gradients: reconstruction updates IDM/FDM, and action regression updates IDM/decoder. The diagram’s summed loss is therefore a schematic of alternating dataset-specific updates. In Stage 2, the padlock marks the pretrained IDM as fixed while it supplies latent targets for the peach policy. The policy receives the current observation; the teacher sees the transition. e03e04e05

What it supports. The latent space is constrained by both prediction and executability. Grounding is learned during representation training, rather than postponed until afterward. At deployment, Algorithm 2 retains the policy and action decoder; the blue forward model does not simulate candidate futures for choosing an action.

Where the evidence stops. The simplified drawing omits the observation history specified in Section IV-A and the loss weight beta. It also is not an inference diagram. Its padlock applies to the pretrained IDM during policy learning, not to every module throughout training.

2. Motivation

2.1 The problem and the proposed response

Source description

How can a robot imitate task-specific demonstrations whose actions are withheld? CLAM separates a large observation corpus, a small action-labeled grounding set, and observation-only expert demonstrations. Experiments isolate action inference using robot data in a single-task, single-embodiment setting. Section III allows some task-specific demonstrations within the labeled mixture; the strongest “no expert labels” wording must therefore be read alongside the purely suboptimal-data ablation. e02e14

2.2 What this reading follows

A video shows what changed, but does not reveal the motor command that caused it. CLAM tackles this missing-label problem in two stages. It first learns continuous transition codes by reconstructing observations while a small labeled dataset teaches the same codes to decode into robot actions. It then labels expert videos with those codes and trains a policy to predict them. The important distinction is where prediction is used: the forward model teaches the representation, while a policy and action decoder execute control. This reading follows the supplied July 2026 v2 through simulated comparisons, physical-robot scores and the ablations that support its design. e01e02e03e04e05e11e12

3. Research context

We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.

Catalog dimensionRecorded classification
Major categoryNot assigned
ArchitectureNot assigned
Prediction paradigmNot assigned
QuadrantNot assigned

This table preserves the labels recorded at reading time. The current major category is Related resources. View the current classification.

3.1 Evidence-based assessment

Insufficient evidence to decide

Reader analysis

The recorded categories are all unassigned, so no existing quadrant claim can be confirmed. Architecturally, CLAM is a multicomponent latent inverse-dynamics pipeline with a training-time forward model and separate deployed policy/decoder. Joint training does not establish a One Model architecture or joint future/action generation at inference. This is an assessment for later classification, not a catalog change. e03e04e05

This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.

4. Problem formulation

4.1 Inputs and outputs

InputsOutputs
  • Observation-only robot corpus D_unlabeled
  • Limited action-labeled transitions D_labeled
  • Task-specific observation-only demonstrations D_unlabeled-expert
  • Reconstructed next observations during pretraining
  • Continuous pseudo-actions for expert trajectories
  • Executable continuous actions from the deployed policy and decoder

4.2 Equations and their role

LCLAM=Lrecon+βLaction-decoder,Lrecon=MSE(o^t+1,ot+1),Laction-decoder=MSE(a^t,at)\mathcal{L}_{\mathrm{CLAM}}=\mathcal{L}_{\mathrm{recon}}+\beta\mathcal{L}_{\mathrm{action\text{-}decoder}},\qquad \mathcal{L}_{\mathrm{recon}}=\operatorname{MSE}(\hat{o}_{t+1},o_{t+1}),\quad \mathcal{L}_{\mathrm{action\text{-}decoder}}=\operatorname{MSE}(\hat{a}_t,a_t)
The source’s unnumbered objective combines next-observation reconstruction with action regression. Hats denote predictions; o is an observation, a an environment action, t the timestep, and beta the loss-balancing hyperparameter. The two terms use unlabeled and labeled batches respectively. e03e04
Lπ=MSE(z^t,zt)\mathcal{L}_{\pi}=\operatorname{MSE}(\hat{z}_t,z_t)
The policy predicts the continuous latent action z_t supplied by the pretrained IDM; the hat denotes its prediction. Expert environment actions are not policy targets. e05

5. Method in detail

5.1 Separate knowing the task from knowing the controls

Source description

CLAM assigns different jobs to its three datasets. The broad observation corpus teaches which transitions can occur. The small labeled dataset connects those transitions to motor commands. The expert observation sequences specify the behavior that the final policy should imitate. This separation explains why task-agnostic play can be useful even when behavior cloning on that play produces a weak task policy: the decoder needs an action mapping, not an expert behavioral distribution. There is nevertheless a source qualification. Section III says that some task-specific demonstrations supplement the labeled mixture, and the user’s expert observation sequences are obtained through robot teleoperation with actions withheld. Thus the practical contribution is reducing expert action-label requirements, with an additional purely suboptimal-data test, rather than demonstrating task learning without expert behavior or robot demonstrations. e02e07e08e14

5.2 Use prediction to learn codes that remain decodable

Reader analysis

The forward model and action decoder impose different demands on the same latent code. Reconstruction encourages the IDM to retain information that predicts the observed change; action regression encourages it to retain information that an observation-independent decoder can turn into a motor command. A reader interpretation is that joint grounding reduces the freedom to choose a reconstruction-friendly code that is difficult to execute. Algorithm 1 supports this interpretation structurally: action-loss gradients reach the IDM, not only the decoder. Table II supplies the behavioral check. Holding joint training off, continuous codes help; holding continuous codes fixed, joint training helps much more. This still does not prove that the latent variables recover a unique physical action representation. The source explicitly describes them as transition-specific and potentially state-dependent. e03e04e11

5.3 Follow the handoff from transition teacher to executing policy

Source description

During expert-data annotation, the IDM can inspect the next observation because the demonstration has already happened. The deployed policy cannot use that future observation; it learns to predict the teacher’s code from the current observation. The learned action decoder then converts its output into an executable command. Algorithm 2 closes the loop by stepping the environment and receiving another observation. The practical policy description adds five-action chunks spanning one second, while leaving refresh and overlap handling unspecified. This handoff is why the forward model’s role should not be described as online planning: it has already been discarded. Evaluation must therefore inspect executed behavior. MetaWorld supplies simulated success rates, and the WidowX table supplies real execution scores with partial credit; neither is merely a video-quality evaluation. e03e05e06e10e12

5.4 Training and inference

During training

Source description

Image CLAM patchifies 64×64 RGB frames into sixteen 16×16 patches, applies spatial then temporal attention, and predicts latent actions from per-frame CLS tokens. The decoder also cross-attends to latent actions. The policy description separately specifies ResNet18 image features and an ACT-like transformer decoder. e06

Source description

Appendix defaults: 500,000 CLAM updates, decoder training every two updates, decoder batch size 128, both loss weights 1, latent dimension 16 and context length 2. Transformer defaults include three encoder/decoder layers, model dimension 256, four heads and dropout 0.1. e16

During inference

Source description

Algorithm 2 repeatedly observes, predicts a latent action, decodes it, and steps the environment. It performs no predicted-video search or FDM rollout. The implementation predicts five-action chunks spanning one second; how overlapping chunks are executed or refreshed is not specified. e05e06

5.5 Implementation flow

  1. Infer and reconstruct transitions

    The latent IDM f_phi observes consecutive observations plus history and produces z_t. The FDM g_psi predicts the next observation from history and z_t. Reconstruction trains a compact, continuous transition code; the source describes this as a low-level, potentially state-dependent action representation, not a skill label. e03

  2. Ground the latent space

    The decoder p_omega maps z_t to the environment action without observing the scene. Reconstruction updates the IDM/FDM; action regression updates both the decoder and IDM on labeled batches. Thus labeled supervision shapes the representation itself. e04

  3. Relabel and imitate

    The pretrained IDM annotates expert observation pairs. A separate policy pi_theta learns their latent targets by mean-squared error. The FDM is discarded; Figure 1 marks the pretrained IDM as locked during policy training. e03e05

6. Experiments & results

CLAM learns continuous action codes from robot observation transitions, grounds them with limited action-labeled data, then imitates expert videos in that latent space. Deployment uses a policy and action decoder; future-observation prediction serves training. Its strongest evidence concerns action-label scarcity within one robot embodiment, with several unresolved protocol details.

6.1 Read the original evidence

Table I. Strong state-based performance spans locomotion and manipulation, with different metrics. Original paper, p. 4 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Start by separating the first two task columns from the next four. HalfCheetah and Hopper report normalized returns; the MetaWorld tasks report success rates. Compare Transformer-CLAM against VPT to evaluate latent relabeling against an inverse model trained only on labeled data. Then compare with BC-Expert, the privileged reference that receives the withheld expert actions. Maroon identifies the strongest nonprivileged method, not the best value including BC-Expert. The asterisks mark the authors’ CLAM variants. Although the source caption calls this a MetaWorld table, the first two columns are DMControl locomotion tasks. Its final average therefore spans heterogeneous metrics. e07e08e09e18

What it supports. Transformer-CLAM obtains 0.72 and 0.81 normalized return on HalfCheetah and Hopper, versus VPT’s 0.32 and 0.41. On Assembly, the success-rate comparison is 0.91 versus 0.40, while BC-Expert reaches 1.00. The gains are substantial within each task, without requiring interpretation of the mixed overall average.

Where the evidence stops. The ± values are preserved exactly, but the paper does not define their statistic or separately specify the state-evaluation rollout count. BC-Expert lacks CLAM’s larger-corpus pretraining, so this comparison does not isolate the causal effect of withholding action labels.

Figure 2. Image-based execution favors CLAM under the paper’s limited-label regime. Original paper, p. 4 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read each group as a separate task before examining the Average group. Orange denotes ST-ViViT CLAM, yellow the privileged BC-Expert reference, and blue VPT. The hatched pink and brown bars identify the discrete-latent baselines. The vertical axis is task success rate, not image-reconstruction quality. Figure 2’s caption specifies 50 evaluation rollouts across three seeds. The plotted legend identifies a single ST-ViViT CLAM series, although the caption names MLP-CLAM and Transformer-CLAM; this edition follows the visible legend and the matching results paragraph. Compare orange and blue within each task to see where the limited labeled-data regime separates these pipelines. e08e10e17e18

What it supports. The accompanying results text reports 76% average success for ST-ViViT-CLAM, versus 9% for LAPO and 20% for LAPA. The plot also shows that approaching the privileged reference is task-dependent. These are simulated policy-execution measurements, providing a different test from learning to reconstruct video frames.

Where the evidence stops. The caption states 100 labeled trajectories, whereas Appendix Table VI lists 50 for MetaWorld. The paper does not reconcile those counts or define the error bars. Retain the displayed comparison without assigning it an invented, fully resolved training protocol.

Table III. Physical execution is measured with partial-credit scores, not binary success counts. Original paper, p. 6 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the scoring definition before interpreting the fractions. Each cell sums ten trials, where partial progress earns 0.5 and full credit earns 1. The caption gives milestone pairs: reach/pick up for block, near/touch for button, touch/close for microwave, and reach/slide for the pot. Therefore 8.5/10 cannot mean eight and a half successful trials. Compare ST-CLAM with VPT to assess the strongest listed nonprivileged competitor, then use BC-Expert as a privileged reference. Figure 3 on page 5 supplies the toy-kitchen scene and camera setup; the table itself records outcomes rather than showing trajectories. e06e08e12

What it supports. ST-CLAM scores 7, 8.5, 8 and 4 points across the four tasks, compared with VPT’s 2.5, 4, 5 and 2. It is close to or above BC-Expert on several scores, but Slide Pot remains the lowest-scoring CLAM task. These observations demonstrate physical deployment under the specified scoring rule.

Where the evidence stops. The original caption says “significantly,” but supplies no significance test or uncertainty interval. Ten aggregate scores do not reveal full-success counts, failure modes or robustness. The caption is retained for its metric definition, not as endorsement of its statistical wording.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
DMControl state-based locomotion

HalfCheetah and Hopper; medium-expert pretraining data and suboptimal labeled trajectories.

Transformer-CLAM: 0.72 ± 0.04 and 0.81 ± 0.05, respectively.

Normalized return

VPT: 0.32 ± 0.04 and 0.41 ± 0.03; BC-Expert: 0.68 ± 0.02 and 0.76 ± 0.04.

These are returns, not success rates. The supplied source does not define the ± statistic or a separate state-evaluation rollout count. e07e08e09

MetaWorld state-based manipulation

Assembly, Bin Picking, Peg Insert and Shelf Place; replay-buffer pretraining and held-out random-medium labeled data.

Transformer-CLAM: 0.91 ± 0.03, 0.82 ± 0.03, 0.79 ± 0.07, 0.93 ± 0.02.

Task success rate

VPT: 0.40 ± 0.08, 0.05 ± 0.02, 0.49 ± 0.06, 0.02 ± 0.00; BC-Expert: 1.00 ± 0.00, 0.94 ± 0.05, 0.91 ± 0.03, 0.93 ± 0.00.

CLAM exceeds VPT across these tasks but does not exceed the privileged reference. Table I’s overall average mixes returns and success rates. e07e08e09

MetaWorld image-based manipulation

Four manipulation tasks; Figure 2 specifies 50 evaluation rollouts across three seeds and 100 labeled trajectories.

ST-ViViT-CLAM: 76%, reported in the results text.

Average task success rate

LAPO: 9%; LAPA: 20%, under the paper’s limited, non-fully-expert labeled-data regime.

This is executed simulated control. The caption and appendix disagree on labeled-data count; the caption also names variants absent from the plotted legend. e08e10e17

MetaWorld continuous-latent and joint-training ablation

Table II, Assembly column; discrete/continuous codes crossed with joint decoder training (JT).

Continuous + JT: 0.69 ± 0.05.

Task success rate

Continuous without JT: 0.28 ± 0.04; discrete without JT: 0.15 ± 0.03; discrete + JT: 0.14 ± 0.04.

Joint grounding helps strongly with continuous codes here. Table II’s caption compares rows 2→3 while changing both factors; rows 1→3 and 3→4 are cleaner contrasts. e11

WidowX physical manipulation

Block, Button, Microwave and Slide Pot; ten trials per task with partial/full credit of 0.5/1.

ST-CLAM: 7/10, 8.5/10, 8/10, 4/10.

Summed task score out of 10

VPT: 2.5/10, 4/10, 5/10, 2/10; BC-Expert: 7.5/10, 8/10, 7.5/10, 2/10.

These totals are not counts of fully successful trials. No per-trial outcomes, uncertainty interval or significance test is supplied. e12

Assembly latent-capacity diagnostic

Figure 4 left: latent dimensions 2, 4, 8 and 16; MetaWorld’s physical action dimension is four.

0.09, 0.11, 0.57 and 0.50, respectively.

Task success rate

Dimension eight exceeds dimension four despite both representing the same control task.

Matching the motor-action dimension is insufficient in this experiment; larger is not monotonically better. e13e16

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Table II. Continuous codes and joint grounding work best together in the reported ablation. Original paper, p. 6 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Treat the row labels as a two-factor design. Disc. and Cont. specify the latent representation; JT means joint action-decoder training, and the negation sign means it is absent. First compare rows one and three to change the representation while holding joint training off. Then compare rows three and four to introduce joint training with continuous codes fixed. Rows one and two test the same grounding change for discrete codes. These comparisons are cleaner than the source caption’s row two-to-three contrast, which changes both factors simultaneously. Read across tasks to check whether the interaction is confined to Assembly. e11

What it supports. On Assembly, moving from discrete to continuous codes without JT changes success from 0.15 to 0.28; adding JT raises the continuous variant to 0.69. Joint training alone changes the discrete variant from 0.15 to 0.14. The combined variant leads across all four reported tasks, supporting the particular combination.

Where the evidence stops. The caption’s two-to-three row comparison does not isolate continuity. These experiments support an interaction in the tested MetaWorld setting, rather than universal necessity for every continuous-control system. The source does not define the ± statistic or fully specify the ablation’s evaluation protocol.

Figure 4. Capacity and demonstration count help, but the scaling variable is inconsistently named. Original paper, p. 6 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. The left panel changes latent dimension while measuring Assembly task success. Its four bars are labeled 0.09, 0.11, 0.57 and 0.50; Appendix Table VII identifies the underlying MetaWorld action dimension as four. The right panel labels increasing trajectory counts and reports 0.39 through 0.73 success. Keep those two interventions separate. The right axis explicitly says D_unlabeled, while its caption describes annotating more trajectories with a pretrained CLAM and page 7’s heading names D_unlabeled-expert. That mismatch prevents a definitive assignment to pretraining-corpus scaling versus downstream demonstration scaling. The categorical bar spacing is not a continuous data-scaling curve. e13e16

What it supports. A latent dimension equal to the motor-action dimension performs poorly here; eight dimensions give substantially higher success, and sixteen do not improve it further. The second panel shows gains with the displayed trajectory count, including smaller gains at the largest counts, but the dataset role needs clarification before replication.

Where the evidence stops. The plot/prose dataset mismatch is unresolved, and no error-bar statistic is defined. The source’s explanation that a fixed action-decoder dataset causes diminishing returns is a hypothesis, not a demonstrated causal diagnosis. This figure does not establish unlimited scaling or a universal optimal dimension.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

Internet-video robustness and human-to-robot transfer remain future work. Expert robot demonstrations are still needed, with labels withheld. The conclusion’s broad “without expert teleoperation” wording exceeds this experimental scope. e02e15

Reader analysis

The appendix lists 50 labeled MetaWorld trajectories versus Figure 2’s 100, and real expert trajectories as “30/50” without task assignment versus approximately 30 per task in Section V. Figure 4’s scaling axis says D_unlabeled, while its caption describes relabeling with a pretrained CLAM and Section VI’s heading says D_unlabeled-expert. These inconsistencies remain unresolved. e07e13e17

Reader analysis

BC-Expert has privileged action labels but lacks CLAM’s larger-corpus pretraining. Near-parity therefore does not isolate the causal value of latent actions or establish that expert labels are unnecessary under matched pretraining. The authors explicitly leave causal attribution open. e08e18

7.2 Questions for discussion

  1. Would continuous codes still help when all baselines receive identical representation pretraining?
  2. How much task-relevant transition coverage must random play provide for reliable grounding?

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

Preserve separate corpus, grounding and expert datasets. Table VI lists 1,000 unlabeled and 20 expert trajectories for each simulated domain; the real setup uses approximately 50,000 unlabeled and 5,000 labeled transitions. The WidowX setup uses external D435 and over-the-shoulder C920 cameras; native action dimensions are 3/6/4/7 for Hopper/HalfCheetah/MetaWorld/WidowX. e07e06e16e17

Reader analysis

The supplied recipe omits optimizer, learning rate, policy-update count, training compute/hardware and software versions. The return cutoff used to select random trajectories is unnamed. Resolve dataset counts, return normalization and encoder initialization before treating a run as an exact reproduction; Section II’s no-off-the-shelf-model wording coexists with Section V’s pretrained ResNet. e03e06e09e14e16e17

8.2 Proposed reproduction checks

The following checks are proposals motivated by the paper. They have not been run as part of this reading.

Check 1: Does grounding help by changing the IDM representation?

Reader-proposed, not performed: on MetaWorld Assembly, compare continuous CLAM with decoder-loss gradients reaching the IDM against a control where that loss updates only the decoder. Keep reconstruction updates, decoder capacity, latent dimension, labeled batches and total updates identical; start from matched initializations. Use the published defaults as a disclosed starting point and evaluate multiple seeds with the same rollout budget. Measure held-out action MSE, next-observation MSE and executed task success. The grounding interpretation predicts improved decoding and success when gradients reach the IDM, even at similar reconstruction error. Comparable outcomes would weaken that interpretation and motivate checking optimization or representation-pretraining confounds. e03e04e10e11e16

Check 2: Can strictly nonexpert play ground expert transitions?

Reader-proposed, not performed: repeat Assembly with provenance-tracked random play only, random/medium data and an expert-labeled reference, using equal labeled-trajectory counts and a fixed observation corpus and expert-video set. Run separate 50- and 100-trajectory conditions because Table VI and Figure 2 disagree. Declare the return cutoff and verify that the random set contains no successful task demonstrations. Compare CLAM with VPT and BC-AL; measure decoding error on held-out expert transitions and policy success using identical seeds and evaluation conditions. If random-only grounding collapses despite matched count, inspect transition coverage before attributing success to expert-label avoidance. This tests a claim motivated by Figure 5 without assuming all main experiments used pure play. e02e07e08e10e14e17

8.3 Reading coverage

Visual audit: All eight original PDF pages were rendered and visually inspected. This includes title/authors/affiliations (p. 1), Figure 1 and data assumptions (p. 2), Algorithms 1–2 and objectives (p. 3), Table I/Figure 2 and datasets (p. 4), camera setup and architectures (p. 5), Tables II–III/Figure 4 (p. 6), Figure 5 and limitations (p. 7), and Appendix Tables IV–VII and references (p. 8). All six final crops were separately viewed. The Table III scoring caption is retained because it defines the indispensable partial-credit metric. Diagram branches, lock marker and loss flow were checked against the caption, objectives and algorithms. No external supplements, code or videos were inspected.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. Appendix coverage: reviewed.

Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.

Text reading scope & known omissions
  • Abstract; I. Introduction (p. 1)
  • II. Related Work (pp. 1–2)
  • III. Problem Setting (pp. 2–3)
  • IV. Continuous Latent Action Models; A. Latent Action Model Training; B. Latent Action Policy Training; Algorithms 1–2 (pp. 3–4)
  • V. Experimental Setup (pp. 4–5)
  • VI. Results (pp. 5–7)
  • VII. Conclusion, including video and embodiment limitations (p. 7)
  • Appendix, Tables IV–VII; References (p. 8)

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Separate supplemental material availability has not been fully verified.
  • The extraction limitation was addressed by separately inspecting every PDF page, Figures 1–5, Tables I–VII and all six final crops. All four supplied text chunks were read completely.
  • Identity notes: the title and seven authors match the catalog. The catalog records a 2025 submission; the inspected artifact is arXiv:2505.04999v2 [cs.RO], dated 30 July 2026. The title page spells Erdem Bıyık with dotless ı; the catalog uses Biyik. Version 1 was not supplied, so scientific differences between revisions were not assessed. The catalog venue reference was not independently verified.
  • Code, project videos and external references were not opened; experiments were not reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e01PDF p. 1, title, author/affiliation block and arXiv marginInspect

Exact title and seven authors; University of Southern California departments; arXiv:2505.04999v2 [cs.RO], 30 July 2026.

Go to primary source ↓
e02PDF pp. 2–3, Section III, Problem SettingInspect

Defines the three datasets, admits some task-specific demonstrations in labeled data, and restricts experiments to action-withheld robot demonstrations in a single-task, single-embodiment setting.

Go to primary source ↓
e03PDF p. 2, Figure 1 and Section II; p. 3, Section IV-A and Algorithm 1Inspect

IDM/FDM reconstruct future observations using continuous latent codes and history. Figure 1 locks the pretrained IDM in Stage 2; Algorithm 1 specifies separate reconstruction and grounding updates.

Go to primary source ↓
e04PDF p. 3, Section IV-A, Latent Action Decoder; Algorithm 1, lines 3–5Inspect

Observation-independent decoder; action loss updates decoder and IDM. Unnumbered CLAM objective combines reconstruction and weighted action MSE.

Go to primary source ↓
e05PDF p. 3, Algorithm 2 and Section IV-B; p. 4, Section IV-BInspect

Discard FDM, annotate expert transitions, train latent-policy MSE, and deploy policy followed by action decoder with environment feedback.

Go to primary source ↓
e06PDF p. 5, CLAM Model Architectures, Latent Policy Architecture and Figure 3Inspect

Space-time patch architecture, CLS latent prediction, ResNet18 policy features, five-action/one-second chunks, and the two named WidowX cameras.

Go to primary source ↓
e07PDF p. 4, Section V, Environments and DatasetsInspect

DMControl medium-expert/suboptimal split; MetaWorld replay-buffer/random-medium split; real approximately 50k/5k transitions and approximately 30 expert demonstrations per task.

Go to primary source ↓
e08PDF p. 5, Baselines; p. 6, Finding 1Inspect

Defines BC-AL, VPT, LAPO, LAPA, DynaMo and privileged BC-Expert; shared components and accessible datasets, with method-specific ability to use them.

Go to primary source ↓
e09PDF p. 4, Table I, all task columns and captionInspect

State-based numerical returns/success rates; Transformer-CLAM, VPT and BC-Expert rows. DMControl uses normalized returns while MetaWorld uses success rates; ± is not defined.

Go to primary source ↓
e10PDF p. 4, Figure 2, plot/legend and caption; p. 6, Finding 2Inspect

Image evaluation uses 50 rollouts across three seeds and states 100 labeled trajectories. Text reports 76% ST-ViViT-CLAM, 9% LAPO, 20% LAPA. Legend identifies ST-ViViT CLAM whereas caption names MLP/Transformer variants.

Go to primary source ↓
e11PDF p. 6, Table II, all rows and caption; Finding 2Inspect

Four-way MetaWorld ablation shows strong continuous-plus-JT gains; caption’s row 2→3 contrast changes both discretization and JT.

Go to primary source ↓
e12PDF p. 6, Table III, all rows and scoring captionInspect

Scores over ten physical trials award 0.5/1 partial/full points. Milestone pairs: block reach/pick up, button near/touch, microwave touch/close, slide pot reach/slide.

Go to primary source ↓
e13PDF p. 6, Figure 4, both panels and caption; p. 7, latent-dimension and scaling paragraphsInspect

Dimension sweep reports 0.09/0.11/0.57/0.50. Data sweep reports 0.39/0.56/0.61/0.69/0.73 for 50/100/150/200/1000 trajectories; axis and prose disagree on which unlabeled dataset varies.

Go to primary source ↓
e14PDF p. 7, Figure 5, Finding 3 and labeled-data paragraphsInspect

Labeled-data quantity and expertise diagnostics; random, random/medium and expert bars read 0.49/0.57/1.00. Random set contains 100 return-filtered trajectories, but no numerical cutoff is given.

Go to primary source ↓
e15PDF p. 7, Section VII, Conclusion, Generalizing to real-world videos and Embodiment gapInspect

Internet-video artifacts and embodiment transfer are future challenges; conclusion uses broad no-expert-teleoperation language.

Go to primary source ↓
e16PDF p. 8, Appendix, Tables IV, V and VIIInspect

Lists pretraining/transformer hyperparameters and native action dimensions; does not provide optimizer, learning rate, training hardware, compute or software versions.

Go to primary source ↓
e17PDF p. 8, Table VI; p. 4, Figure 2 caption and Section VInspect

Trajectory split table lists simulated 1000/50/20 and real approximately 500/50/30-or-50; main text/caption use different labeled/expert counts without reconciliation.

Go to primary source ↓
e18PDF p. 6, Finding 1, first two paragraphsInspect

Authors attribute BC-Expert parity to possible representation/pretraining advantages and latent regularization, explicitly leaving full causal attribution to targeted ablations.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.