BridgeData V2: A Dataset for Robot Learning at Scale
1. Paper overview
In one sentence: A shared manipulation dataset supports several action-learning methods and cross-lab reuse, while its single robot type, uneven coverage and limited evaluations bound the generalization claim. e02e03e05e06e07e11e14e15e16e17
| At a glance | What to know |
|---|---|
| Research problem | Source description Robot datasets tied to a few scenes are difficult for another laboratory to reuse. BridgeData V2 targets broad task and environment coverage while making task instructions necessary: multiple feasible behaviors share a scene, so the initial image should not uniquely reveal the intended task. e02e03 |
| Core mechanism | Source description The release is described as 60,096 trajectories: 50,365 expert demonstrations and 9,731 scripted trajectories, covering 13 skills, 24 environments and more than 100 objects. A skill groups similar motions; a task groups similar language instructions. Figure 3 shows substantial concentration in object manipulation and toy kitchens. e01e04 |
| A key reported result | Unseen pick-and-place with broader training skill coverage: 0.65 with 13 skills. Task success rate, higher is better. GCBC trained on approximately matched dataset sizes: 3 skills with 28k trajectories versus 13 skills with 27k; 20 trials per policy. 0.30 with 3 skills; reader-calculated difference of 35 percentage points. Supports positive transfer in this test. The paper does not establish an object/environment-matched diversity control or report seed-level uncertainty. e16 |
| Reading caution | Source description Tasks are generally low-precision and omit demanding force control, dynamic manipulation and tight-tolerance insertion. Collection occurred at one institution on one robot type; morphology generalization is future work. e17 |
Core contributions
- Source description
The release is described as 60,096 trajectories: 50,365 expert demonstrations and 9,731 scripted trajectories, covering 13 skills, 24 environments and more than 100 objects. A skill groups similar motions; a task groups similar language instructions. Figure 3 shows substantial concentration in object manipulation and toy kitchens. e01e04
- Source description
The study evaluates four goal-conditioned and two language-conditioned methods, including imitation learning and offline reinforcement learning. It tests seen tasks, novel objects/environments, another institution, and model/data scaling. These are demonstrations of dataset utility, not isolated comparisons of every architectural decision. e07e10e11e12e16
Figure 2. The collection system makes the robot reproducible while allowing scene and camera variation. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the black callout identifying the WidowX 250 arm. The red label marks the fixed over-the-shoulder camera with depth, while the two green callouts identify cameras whose poses change during collection. These strokes identify physical equipment; they are not a computational graph. The caption additionally describes a wrist-mounted RGB camera, although it is not separately labeled in this photograph. Images are saved at 640×480 and control runs at 5 Hz. Appendix C identifies the specific cameras and VR headset. During policy evaluation, Section 4 uses only the fixed camera’s RGB view. e03e05e06e07
What it supports. The source describes an approximately $4,000 setup built from publicly available parts at publication. Its methodological role is to let laboratories share one action interface while collecting variation in scenes, objects and viewpoints. This supports a practical reuse experiment without requiring each recipient to reproduce one exact kitchen arrangement.
Where the evidence stops. The photograph shows the developed setup, not a guarantee that every trajectory contains every sensor. Figure 6 documents missing camera modalities across much of the dataset, and all reported baselines use a single RGB view.
2. Motivation
2.1 The problem and the proposed response
Robot datasets tied to a few scenes are difficult for another laboratory to reuse. BridgeData V2 targets broad task and environment coverage while making task instructions necessary: multiple feasible behaviors share a scene, so the initial image should not uniquely reveal the intended task. e02e03
2.2 What this reading follows
Imagine acquiring a robot and training a useful policy before collecting demonstrations in your own laboratory. BridgeData V2 asks how far a diverse shared dataset can make that practical. Its collection protocol combines many feasible tasks in each scene, varied objects and camera arrangements, and language descriptions added after collection. The experiments then train six goal- or language-conditioned policies and execute them on physical manipulation tasks. This reading follows the collection hardware, the available sensor coverage, and three distinct meanings of generalization: unfamiliar objects or scenes, another laboratory, and benefits from broader training skills. The reported gains are useful evidence, but neither universal task competence nor robot-morphology transfer is established. e02e03e05e06e07e11e14e15e16e17
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 dimension | Recorded classification |
|---|---|
| Major category | Datasets |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The dataset category and Not applicable architecture, prediction paradigm and quadrant are supported: this resource benchmarks separate action policies and introduces no joint future/action predictor or inverse-dynamics world-action architecture. Multitask demonstrations are supported; the combined Cross-robot & multitask subcategory must not be read as evidence of multiple robot morphologies. Cross-lab transfer uses the same robot type. e07e09e10e11e15e17
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
| Inputs | Outputs |
|---|---|
|
|
5. Method in detail
5.1 Make the task specification informative during collection
The collection protocol tackles a subtle shortcut: if every scene has only one demonstrated task, a policy can infer what to do from the scene and ignore the instruction. BridgeData V2 instead places several usable objects and affordances together and lets collectors perform any feasible task. They do not have to reset the scene after each trajectory, and camera, object and workspace randomization occurs every fifty trajectories. Language descriptions are added afterward, emphasizing the final location of moved objects. For goal-conditioned learning, the training goal comes from a later image in the same trajectory. Reader interpretation: these two conditioning routes reuse the same behavioral experience while giving the learner a reason to distinguish different intended outcomes within a shared scene. The paper does not directly measure how completely this removes instruction-ignoring shortcuts. e02e03e08
Figure 6. Sensor availability varies across trajectories because cameras were added during collection. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each wedge as a combination of available sensors, not a skill or a success rate. Purple denotes only the primary over-the-shoulder view. Red adds the randomized alternative views; green also includes depth, orange also includes the wrist camera, and blue includes both depth and wrist. The caption says depth shares the primary camera’s perspective and that additional cameras were introduced as collection progressed. Thus camera availability records part of the collection history. Keep that distinction in mind when comparing a model trained on all primary-view data with one requiring multiple views. e01e05e06e07
What it supports. The primary-view-only wedge is the largest, and the all-modality subset is much smaller. The six policy baselines use the primary RGB view, so their comparison does not test whether depth or extra viewpoints improve control. A multimodal experiment would first need to account for which trajectories remain eligible.
Where the evidence stops. The printed counts sum to 60,151, whereas the abstract states 60,096 trajectories: a reader-calculated discrepancy of 55. The faithful crop preserves those labels. No supplied explanation reconciles the totals, so avoid deriving exact release percentages from them.
5.2 Follow the prediction target all the way to robot action
The evaluated systems predict actions rather than future images. GCBC combines current and goal images before a ResNet and action head. D-GCBC replaces the simple action prediction with a diffusion distribution trained using a DDPM objective, without learning a value function. ACT changes the temporal unit to a five-action sequence and uses a conditional VAE with transformers. CRL instead learns goal-conditioned value representations with a temporal-difference contrastive objective and an imitation regularizer. The language alternatives inject a sentence embedding through FiLM or use RT-1’s tokenized history and action prediction. These choices affect the commands executed by the same robot interface. The authors report that D-GCBC sometimes oscillates between action modes, whereas ACT has action chunks and RT-1 has history. That behavioral explanation is a qualitative observation, not an isolated temporal-consistency ablation. e07e08e09e10e11e12
5.3 Separate reuse evidence from explanations of why it works
The experimental sequence answers progressively different questions. Seen tasks test whether the shared data can support useful policies despite changed object positions and distractors. The unseen suite changes objects, environments or both, and its footnotes are essential because those conditions are not interchangeable. Cross-lab evaluation then adds an institutional domain shift without collecting local training data. None of these comparisons alone identifies which dataset property causes the success. Figure 5 comes closer by approximately matching trajectory counts while changing the number of skills; the broader set improves one unseen pick-and-place task. Reader interpretation: this motivates a positive-transfer hypothesis, but object and environment coverage still need tighter controls. Appendix D provides complementary evidence that the expanded release is useful, while jointly changing size and diversity and leaving a printed-average discrepancy unresolved. e12e13e14e15e16e18
5.4 Training and inference
During training
Goal-conditioned methods uniformly sample a future observation in the trajectory as the training goal. GCBC and D-GCBC use random crops, resizing and color jitter. D-GCBC uses the DDPM objective; ACT uses a conditional VAE objective. CRL retains a temporal-difference contrastive objective with GCBC regularization coefficient 0.2. e08e09e10
GCBC, D-GCBC, CRL and LCBC specify Adam at 0.0003 with 2,000 warmup steps. ACT reports three days on one Nvidia 2080Ti. RT-1 increases transformer sequence length from 6 to 15 and tokenizes actions scaled to [−1, 1] using 256 bins. These details do not specify a matched compute budget. e08e09e10e11
During inference
Policies transform camera observations and task conditioning into robot commands at the setup’s 5 Hz control frequency. ACT predicts action sequences; RT-1 incorporates history. The evaluations concern physical task execution, not generated-video quality. Exact ACT chunk execution/aggregation settings and diffusion sampling schedules are not restated here. e05e07e09e11e12
The second institution evaluates zero-shot, without collecting new training data there. The authors’ expectation that local fine-tuning would improve performance is a proposal, not a measured result in this paper. e15
5.5 Implementation flow
- Collect varied tasks without mandatory resets
Collectors teleoperate a WidowX 250 using VR and perform feasible tasks in shared scenes. Object and robot resets are not required between trajectories. Every 50 trajectories, camera poses, objects and workspace position are varied. A randomized pick-and-place script adds frequent failures; annotators later describe performed tasks, emphasizing moved objects’ final locations. e03e05
- Expose a common policy interface
All evaluated policies use only the fixed over-the-shoulder RGB view. Images are 128×128 except RT-1 at 320×256. Actions have six continuous relative Cartesian pose components and a discrete gripper component. Goal images or language specify the task; the dataset is not itself a learned dynamics model. e07e11
- Compare distinct action learners
GCBC stacks observation and goal channels, then uses ResNet-34 and three 256-unit layers to predict actions. D-GCBC conditions an action diffusion model on the same type of encoding, without learning a value function. ACT uses ResNet-18 with a transformer conditional VAE to predict five-action chunks. CRL learns goal-conditioned values with contrastive representations and a shared policy/value ResNet-34 backbone. e08e09e10
- Condition action prediction on language
LCBC passes frozen MUSE language embeddings through two fully connected layers and applies FiLM at each ResNet-34 block before action prediction. RT-1 uses pretrained image/language tokenization and a decoder-only transformer for discretized actions; it alone uses observation histories in this comparison. e11
6. Experiments & results
BridgeData V2 makes varied manipulation experience reusable through a common robot platform, goal images and language labels. Six offline policy baselines demonstrate useful but uneven generalization. Its strongest controlled lesson is that broader skill coverage can help an unseen pick-and-place task; cross-morphology transfer remains untested.
This dataset paper supplies a collection-system photograph rather than a unified neural architecture diagram. Figure 2 therefore serves as the method visual; Section 4 and Appendix B describe the separate evaluated policies. The edition includes original quantitative results and both scaling and earlier-dataset comparisons, without inventing a world-model architecture. e05e07e08e09e10e11
6.1 Read the original evidence
Table 3. Novel objects and novel environments expose different strengths in goal and language conditioning. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the footnotes: a star marks unseen objects in a seen environment, a dagger marks unseen objects and an unseen environment, and a double dagger marks seen objects in an unseen environment. Then compare methods within a row. GCBC, D-GCBC, ACT and CRL receive goal images; LCBC and RT-1 receive language. Each task-method entry summarizes ten physical trials. The bottom row averages different tasks and novelty conditions, so it compresses substantial variation. In particular, marker-in-bowl asks for both object and environment transfer, while spoon-on-cloth tests a new environment with familiar objects. e07e10e11e12e14
What it supports. GCBC averages 0.60 and RT-1 0.50, but neither wins every task. Marker-in-bowl gives GCBC 0.6 and CRL 0.7 while both language methods score 0.0; RT-1 instead reaches 1.0 on spoon-on-cloth. The source associates unseen-object language difficulties with object names not grounded in training data.
Where the evidence stops. These are small samples without reported confidence intervals. Conditioning, image resolution, history and architecture differ across methods. The table supports task-specific transfer findings, not a controlled claim that one conditioning modality is intrinsically superior.
Table 4. The same dataset enables zero-shot execution at a receiving institution, with uneven changes in success. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The arrow in each cell means Lab 1 to Lab 2, not a training update or a guaranteed improvement. Compare the left and right values for the same method and task before reading the average row. The three tasks match training-seen tasks, but the receiving institution changes robot setup, camera placement, lighting and object instances. Section 5.2 says this is zero-shot evaluation with no new training data collected in Lab 2. The table uses ten trials per task. A nonzero average means a method succeeds somewhere in this small suite, not on every task. e05e15e17
What it supports. RT-1 changes from 0.47 to 0.40 on average, while GCBC changes from 0.30 to 0.13. All methods retain nonzero Lab 2 averages. ACT and CRL actually increase their averages, so the table supports uneven domain transfer rather than uniform performance loss across algorithms.
Where the evidence stops. Only one receiving institution and three tasks are evaluated. The setup retains the robot type, so this is not morphology transfer. Expected improvement from local fine-tuning is discussed but not measured here.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Eight seen manipulation tasks Training-seen tasks with varied object positions, distractors and lighting; 10 physical trials per task and method. | GCBC 0.49; D-GCBC 0.49; ACT 0.41; CRL 0.42; LCBC 0.23; RT-1 0.49. Mean task success rate, higher is better | RT-1 versus LCBC: 0.49 versus 0.23. RT-1 improves over the simpler language baseline, but image resolution, history and action representation also differ. No uncertainty intervals are supplied. e07e11e12e13 |
| Six tasks with unseen objects and/or environments Table 3 separates unseen objects, unseen environments, and both; 10 trials per task and method. | GCBC 0.60; D-GCBC 0.55; ACT 0.28; CRL 0.52; LCBC 0.08; RT-1 0.50. Mean task success rate, higher is better | On marker-in-bowl with unseen objects and environment, GCBC/D-GCBC achieve 0.6, CRL 0.7, and both language methods 0.0. Generalization depends strongly on task and conditioning. The unseen suite differs from the seen suite, so its higher GCBC average does not establish that novelty helps. e12e13e14 |
| Zero-shot transfer to another institution Three matched tasks in Lab 1 and Lab 2, with changed setup, camera placement, lighting and object instances; 10 trials per task. | RT-1 0.47 → 0.40; GCBC 0.30 → 0.13. Mean success rate, Lab 1 → Lab 2 | ACT improves 0.03 → 0.10 and CRL 0.13 → 0.20; all six methods have nonzero Lab 2 averages. Transfer is feasible but uneven; the table does not show universal degradation. Only three tasks and one receiving institution are tested. e15 |
| Unseen pick-and-place with broader training skill coverage GCBC trained on approximately matched dataset sizes: 3 skills with 28k trajectories versus 13 skills with 27k; 20 trials per policy. | 0.65 with 13 skills. Task success rate, higher is better | 0.30 with 3 skills; reader-calculated difference of 35 percentage points. Supports positive transfer in this test. The paper does not establish an object/environment-matched diversity control or report seed-level uncertainty. e16 |
| Unseen pick-and-place after expanding the earlier dataset Appendix D: GCBC on BridgeData V1 + PTR (13k trajectories, 15 environments, 11 skills) versus V2 (60k, 24, 13); trial count not separately specified. | V2: marker-in-bowl 0.65; mushroom-in-pot 0.70. Per-task success rate, higher is better | V1 + PTR: 0.05 and 0.10. Size and diversity change together. Table 6 prints V2 average 0.70, whereas the displayed rows average 0.675; preserve the discrepancy. Its V2 marker result also differs from Table 3’s 0.60, so the evaluations are not pooled. e18e14 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 5. Capacity, trajectory count and skill diversity are probed through separate GCBC comparisons. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the three panels as separate experiments. The left plot changes the image encoder among ResNet-18, ResNet-34 and ResNet-50 for moving a spoon. The middle plot changes the retained dataset fraction, with blue circles for a seen task and orange squares for an unseen task. The text says subsampling is stratified by task to preserve diversity. The right table instead changes skill coverage from three skills to thirteen while holding trajectory counts approximately constant, 28k versus 27k. Its target is an unseen pick-and-place task. The caption specifies twenty trials per policy; no error bars appear. e16
What it supports. Broader skill coverage raises reported success from 0.30 to 0.65, a reader-calculated gain of 35 percentage points. This is evidence that other skills can help the tested pick-and-place behavior. The middle plot also shows a boundary: unseen-task performance plateaus from one-half of the data onward.
Where the evidence stops. The skill comparison approximately controls count but does not establish matched object and environment distributions. Results cover a narrow task selection without seed-level uncertainty. Positive slopes in these plots are not a general scaling law.
Table 6. The expanded dataset improves both displayed tasks, but the aggregate contains an unresolved arithmetic mismatch. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the two task rows before the bold average. The dagger denotes unseen objects and environment for marker-in-bowl; the double dagger denotes seen objects in an unseen environment for mushroom-in-pot. Appendix D compares GCBC trained on original BridgeData plus PTR data against GCBC trained on BridgeData V2. The caption describes changes from 13k to 60k trajectories, 15 to 24 environments, and 11 to 13 skills. This is an expansion comparison that changes several factors together, unlike the approximately count-matched skill experiment in Figure 5. Appendix D does not separately state a trial count. e14e16e18
What it supports. Both displayed task values improve: marker-in-bowl rises from 0.05 to 0.65 and mushroom-in-pot from 0.10 to 0.70. However, the two V2 entries average 0.675 by reader arithmetic, while the table prints 0.70. The crop and report preserve that discrepancy rather than silently correcting the published aggregate.
Where the evidence stops. The source supplies no explanation for the average mismatch. The V2 marker result also differs from Table 3’s 0.60. Keep these evaluations separate, and do not attribute the gain uniquely to dataset size or skill diversity.
7. Analysis & limitations
7.1 What the evidence leaves open
Tasks are generally low-precision and omit demanding force control, dynamic manipulation and tight-tolerance insertion. Collection occurred at one institution on one robot type; morphology generalization is future work. e17
Sensor coverage is incomplete. Figure 6 labels total 60,151 trajectories, 55 above the stated 60,096. Figure 3’s category counts likewise total 50,420 rather than 50,365 demonstrations. These unreconciled counts prevent treating the diagrams as an exact release inventory. e01e04e06
The small evaluation samples and multi-factor baseline differences limit causal ranking. Unseen object names challenge language methods. The paper does not provide a named data license, exact train/validation file split, or task-level success adjudication rules. e12e14e19
7.2 Questions for discussion
- Would the skill-diversity gain survive exact matching of objects, environments and trajectory counts?
- How much of the unseen-object language gap comes from vocabulary grounding versus perception and control?
8. Reproducibility audit
8.1 Requirements and known gaps
Match the WidowX 250, 5 Hz control and primary camera view. Appendix C identifies RealSense D435, two Logitech C920 webcams, a wrist Raspberry Pi camera and Meta Quest 2 teleoperation. Reported setup cost is approximately $4,000 at publication, not a current quote. e05
A faithful rerun needs dataset/subset identifiers, scripted-data inclusion per baseline, training steps, numeric batch sizes, seeds, checkpoint selection and omitted inference settings. Appendix B gives useful adaptations but delegates other ACT/RT-1 hyperparameters to prior papers. Public data/models are author release claims; the supplied PDF does not verify present access or licensing. e08e09e11e19
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: Test skill transfer under matched object and environment coverage
Reader-proposed check, not performed: reproduce the three-skill versus thirteen-skill GCBC comparison using identical trajectory counts, encoder, optimizer budget, observation view and scripted-data inclusion. Match the available object/environment strata as closely as the data permit, report unmatched strata explicitly, and repeat subset sampling and policy training across seeds. Evaluate the same held-out pick-and-place task with paired initial-state distributions and report trial counts plus uncertainty. Retention of a positive gap after these controls would strengthen the skill-transfer explanation; disappearance or reversal would indicate that other coverage differences contributed to the original 0.30 versus 0.65 result. e03e07e08e16e19
Check 2: Distinguish camera benefits from trajectory-selection effects
Reader-proposed check, not performed: first audit dataset trajectory identifiers and sensor manifests against Figure 6’s counts and the stated release total, recording any unresolved mismatch. Then fix one architecture with optional camera inputs and train a primary-RGB-only version and a multiple-view version on exactly the same eligible trajectories. Keep goals, actions, augmentations, parameter budget and training updates matched; repeat seeds and evaluate identical seen and shifted-camera tasks. A third primary-view policy trained on the larger primary-view pool tests the cost of discarding trajectories. Gains only against that third policy, but not within the paired subset, would implicate data selection rather than additional sensing. e01e03e05e06e07e08e12e19
8.3 Reading coverage
Visual audit: Visually inspected the title/author page, all seven figures and all six tables, plus every page supporting method, training, evaluation, hardware and reproduction statements. Cross-checked Figure 2 callouts against its caption and Appendix C, Table 3/6 novelty markers against footnotes, Table 4 arrow direction against Section 5.2, and Figure 5 axes/legend against Section 5.3. All six final original crops were viewed; the Table 6 crop was extended to retain the complete footnote. Figure 3/6 count mismatches and Table 6’s average mismatch remain explicit. All five text chunks, including reference pages 9–11, were read; those reference pages were not visually inspected. External videos, dataset files, code, model weights and separate supplements remain outside this review.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14. 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
- Title, authors, affiliations and abstract (p. 1)
- 1 Introduction (p. 2)
- 2 Related Work (pp. 2–3)
- 3 BridgeData V2: system, collection and composition (pp. 3–5)
- 4 Offline Learning Methods, 4.1–4.2 (pp. 5–6)
- 5 Experiments, 5.1–5.3 (pp. 6–8)
- 6 Discussion, Limitations, and Future Work (p. 8)
- Acknowledgements and References (pp. 9–11)
- A Data Statistics (p. 12)
- B Learning Method Implementation Details, B.1–B.6 (pp. 12–14)
- C Hardware Setup and D Comparison to the Original Bridge Data (p. 14)
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 retained PDF was visually inspected on pages 1–8 and 12–14, including all figures and tables; the extraction-only image limitation was addressed. Reference pages 9–11 were read as text.
- No separate supplements were supplied. The linked website, videos, dataset files, code and checkpoints were not inspected; no experiments were reproduced.
- Identity/version: the title page identifies CoRL 2023 and matches the catalog title and main author string. The catalog BibTeX uses expanded name forms and a different author order; metadata follows the printed title-page credits. No separate revision identifier/date is printed, and no alternate edition was supplied or compared.
- No formal equations are presented in the supplied paper; objectives are described in prose and equations are therefore left empty.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author block, abstract and Figure 1 caption
The printed title and fourteen authors match the catalog’s main author string. Affiliations are UC Berkeley, Stanford, Google DeepMind and CMU. The page identifies CoRL 2023 and states 60,096 trajectories, 24 environments and natural-language labels for all data.
Go to primary source ↓e02PDF p. 2, Section 1, paragraphs 2–6
The motivation is reusable data across laboratory scenes, flexible goal/language conditioning, and multiple feasible tasks per environment; the contribution combines a dataset and six-method evaluation.
Go to primary source ↓e03PDF p. 4, Section 3.2, all paragraphs
No mandatory between-trajectory resets; randomize cameras, objects and workspace every 50 trajectories. Scripted pick-and-place frequently fails. Post-hoc annotators describe performed tasks and final object locations.
Go to primary source ↓e04PDF pp. 4–5, Section 3.3 and Figure 3; p. 12, Table 5
The text defines skill versus task and gives 50,365 demonstrations, 9,731 scripted trajectories, 13 skills, 24 environments and over 100 objects. Figure 3 is demonstration composition, dominated by object manipulation and toy kitchens. Reader arithmetic: both panels’ labeled category counts sum to 50,420, inconsistent with the text’s demonstration count.
Go to primary source ↓e05PDF p. 3, Section 3.1; p. 4, Figure 2 and caption; p. 14, Appendix C
WidowX 250 setup costs about $4,000 at publication. Images are saved at 640×480 and control is 5 Hz. Fixed RGBD, two randomized RGB and wrist RGB cameras are described; Appendix C names D435, C920, Raspberry Pi camera and Meta Quest 2.
Go to primary source ↓e06PDF p. 12, Figure 6 labels and caption
The entire-dataset camera chart labels 33,245 primary-only, 13,239 primary plus randomized, 5,819 adding depth, 5,533 adding wrist, and 2,315 with all modalities. Sensors were added during collection. Reader arithmetic sums the labels to 60,151, 55 more than the abstract total.
Go to primary source ↓e07PDF p. 5, Section 4 opening and Section 4.1
Policies use only over-the-shoulder RGB, 128×128 except RT-1 at 320×256. Actions comprise 6D relative Cartesian end-effector motion plus discrete gripper control. GCBC, D-GCBC and ACT are goal-conditioned action learners.
Go to primary source ↓e08PDF p. 12, Appendix B closing paragraph; p. 13, Appendices B.1–B.2
Future trajectory observations are sampled uniformly as goals. GCBC uses channel stacking, ResNet-34, three 256-unit layers, crop/resize/color augmentation and Adam 3e-4 with 2,000 warmup steps. D-GCBC uses analogous encoding, augmentation and optimizer with DDPM action diffusion, without a value function.
Go to primary source ↓e09PDF p. 5, Section 4.1 ACT paragraph; p. 13, Appendix B.3
ACT uses a transformer conditional VAE. The adaptation stacks observation/goal inputs to ResNet-18, reduces chunk size to 5 for 5 Hz and roughly 50–100-step trajectories, and reports training for three days on one Nvidia 2080Ti; other hyperparameters refer to the original ACT paper.
Go to primary source ↓e10PDF p. 6, Section 4.1 CRL paragraph; p. 13, Appendix B.4
CRL learns contrastive goal-conditioned value representations, shares ResNet-34 between policy and value, uses TD-style training and GCBC regularization 0.2, and specifies Adam 3e-4 with 2,000 warmup steps. Numeric batch size is not given.
Go to primary source ↓e11PDF p. 6, Section 4.2; p. 14, Appendices B.5–B.6
LCBC uses frozen MUSE, two language layers, FiLM throughout ResNet-34 and an action head; Adam 3e-4 and 2,000-step warmup. RT-1 uses pretrained tokenization, a decoder-only transformer and observation history; sequence length changes 6 to 15 and actions are scaled to [−1,1] with vocabulary 256.
Go to primary source ↓e12PDF p. 6, Section 5 and Section 5.1; p. 13, Figure 7 and caption
The study targets dataset utility rather than rigorous isolation of all method choices. Seen tasks vary positions, distractors and lighting; ten trials are collected per task/method. Figure 7 illustrates seen/unseen tasks. Section 5.1 reports pauses and D-GCBC oscillations qualitatively and discusses unseen-name grounding difficulties.
Go to primary source ↓e13PDF p. 7, Table 2, Average row and caption
Seen-task averages are GCBC .49, D-GCBC .49, ACT .41, CRL .42, LCBC .23 and RT-1 .49 across eight tasks, with ten trials per task. No uncertainty intervals appear.
Go to primary source ↓e14PDF p. 7, Table 3, all rows, caption and novelty footnotes
Unseen-suite averages are .60/.55/.28/.52/.08/.50 for GCBC/D-GCBC/ACT/CRL/LCBC/RT-1. Marker-in-bowl is .6/.6/.2/.7/.0/.0. Star indicates unseen objects only, dagger both objects/environment unseen, and double dagger unseen environment only; ten trials per task.
Go to primary source ↓e15PDF pp. 6–7, Section 5.2; p. 7, Table 4, arrows, rows and caption
Three tasks are evaluated zero-shot at another institution without new local training data. Lab 1→2 averages are GCBC .30→.13, D-GCBC .23→.13, ACT .03→.10, CRL .13→.20, LCBC .13→.03, RT-1 .47→.40. Ten trials per task; fine-tuning benefit is anticipated rather than measured.
Go to primary source ↓e16PDF pp. 7–8, Section 5.3; p. 8, Figure 5, axes, legend, table and caption
GCBC capacity is varied on spoon moving; task-stratified data subsampling preserves task diversity. Plots show capacity gains and a plateau for unseen-task success after half the dataset. Skill comparison uses three skills versus thirteen, approximately 28k versus 27k trajectories, and .30 versus .65 success on an unseen pick-and-place task, with twenty trials per policy.
Go to primary source ↓e17PDF p. 8, Section 6, limitations paragraph
The authors identify low-precision tasks, lack of complex force/dynamic manipulation, single-institution collection and difficulty standardizing on one robot. Multi-robot morphology generalization is proposed as future data collection.
Go to primary source ↓e18PDF p. 14, Table 6, rows, footnotes and caption; Appendix D
GCBC compares V1+PTR (13k trajectories/15 environments/11 skills) to V2 (60k/24/13). Marker values are .05 versus .65 and mushroom values .10 versus .70. Printed averages are .08 versus .70. Reader arithmetic yields .675 for the two V2 rows; no separate trial count is stated.
Go to primary source ↓e19PDF p. 1, abstract; p. 3, Table 1; pp. 4–8, Sections 3–6; pp. 12–14, Appendices A–D
Authors describe public data and pretrained models. The supplied paper does not state a named dataset license, exact file-level train/validation partitions, or detailed task-success adjudication. Appendix B supplies selected implementation settings but not all methods’ steps, numeric batch sizes, seeds, checkpoint selection or per-baseline scripted-data inclusion.
Go to primary source ↓8.5 Primary sources
BridgeData V2: A Dataset for Robot Learning at Scale ↗
PDF · 7,903 extracted words
Source fingerprint
52dc04866e173336d090f0b42b373d4b495a4a35cedc4f756b4d064936383825