GR00T N1: An Open Foundation Model for Generalist Humanoid Robots
1. Paper overview
In one sentence: GR00T N1 turns heterogeneous robot and video experience into an embodiment-aware action policy, improving tabletop adaptation while leaving broad humanoid autonomy untested. e03e04e05e07e13e14e15e18
| At a glance | What to know |
|---|---|
| Research problem | Source description Humanoid demonstrations are expensive, while robot datasets differ in sensors, control conventions and action dimensions. GR00T N1 addresses how to share visual and behavioral knowledge across these incompatible embodiments and action-less videos, then adapt with limited target-robot demonstrations. Its data pyramid combines broad web/video priors with synthetic trajectories and robot-specific grounding. e02e05e07 |
| Core mechanism | |
| A key reported result | GR-1 real-world tabletop manipulation: Full data: 76.8%; 10% data: 42.6%. Reported average policy score (%) under the paper's partial-credit protocol. Post-training on full versus 10% task demonstrations; pick-and-place, articulated, industrial and coordination categories. Usually 10 trials per task with partial credit; machinery packing uses five trials and object counts within 30 seconds. Diffusion Policy: 46.4% full and 10.2% low-data; differences are 30.4 and 32.4 percentage points. Evidence for adaptation efficiency, not binary completion on every episode. Low-data GR00T remains 3.8 points below full-data Diffusion Policy. Baselines do not share GR00T's pre-training. e12e14 |
| Reading caution | Source description The authors limit current scope to short-horizon tabletop manipulation and identify physical consistency and diversity of synthetic trajectories as unresolved challenges. Section 4.5 also describes a handover skill lost after right-hand-only post-training. e18 |
Core contributions
Figure 3. Shared visual-language conditioning meets embodiment-specific motor interfaces. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Trace three inputs separately. Image and language tokens pass through Eagle-2; its outgoing features connect to the cross-attention layers. Robot state and noised action chunks pass through green, embodiment-specific encoders before the alternating attention stack. The green output decoder returns actions in the corresponding robot's space. The dotted return path denotes repeated denoising of a candidate chunk, not a trajectory of measured environmental feedback. Section 2.1 specifies a 16-action horizon and four sampling iterations. Read the snowflake with Section 2.3 and Table 6: the language component stays frozen, while the separate vision encoder is trainable. e03e04e07
What it supports. The coupling is visible in the cross-attention arrows, and the embodiment dependence is visible in the green MLPs. This supports the catalog's dual-system VLA classification: language-conditioned perception supplies features to a motor generator. A jointly trained system can still contain distinct modules with different functions and interfaces.
Where the evidence stops. The diagram does not resolve the printed flow convention. Page 5 trains against epsilon minus the action chunk but advances the noise-to-action interpolation by adding that field. The report preserves this sign inconsistency instead of treating the loop as executable pseudocode.
2. Motivation
2.1 The problem and the proposed response
Humanoid demonstrations are expensive, while robot datasets differ in sensors, control conventions and action dimensions. GR00T N1 addresses how to share visual and behavioral knowledge across these incompatible embodiments and action-less videos, then adapt with limited target-robot demonstrations. Its data pyramid combines broad web/video priors with synthetic trajectories and robot-specific grounding. e02e05e07
2.2 What this reading follows
A robot video contains evidence about motion but usually lacks the motor commands needed to imitate it. GR00T N1 tackles this mismatch with two complementary choices: a vision-language backbone conditions a separate action transformer, and video-derived labels make diverse experience usable during training. This reading follows the route from pixels and language to executable action chunks, then examines what the reported experiments establish. The strongest findings concern post-training on simulated manipulation and physical GR-1 tabletop tasks. Two source inconsistencies matter for reproduction: the printed flow target and update use incompatible signs, and the main DexMimicGen summary differs from the appendix. e03e04e05e07e13e14e15e18
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 | VLA |
| Architecture | Dual-system |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded VLA and Dual-system labels match the separate Eagle-2 and action DiT connected by cross-attention. The policy predicts actions from current observations; generated futures and inverse dynamics supply offline supervision. It is therefore outside the catalog's joint future/action versus inverse-dynamics world-action quadrants. Joint optimization alone would not justify a One Model label, and the auxiliary detection loss is not an inference-time controller. e03e04e05e07e08e09
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 |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Follow conditioning into an action distribution
Start with a language instruction and the robot's current views. Eagle-2 supplies visual-language embeddings; it does not need to emit a textual plan before movement. The action pathway receives those embeddings through cross-attention, together with proprioceptive state and a noisy candidate chunk. Embodiment-specific encoders and decoders allow different robot state/action dimensions to connect to a shared transformer. Training teaches a conditional vector field; inference uses repeated updates from Gaussian noise to obtain a chunk of motor commands. The frozen-language detail matters: the paper's end-to-end description does not mean every component changes during optimization. The vision encoder, adapters and DiT can adapt while the language component remains fixed. The printed sampling equations require a sign clarification before a faithful implementation can be claimed. e03e04e07
Figure 4. Retrieved frame pairs illustrate the intended motion semantics of latent actions. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each small horizontal pair shows a current and future frame from one embodiment; compare the displacement within a pair before comparing across rows. The two large groups are labeled by the authors as opposite directions of right-arm movement. Their visual backgrounds, manipulators and camera perspectives differ. The training mechanism is described on page 5: a VQ-VAE encodes the frame pair, and a decoder reconstructs the future using the current frame plus the latent. After training, the continuous embedding before quantization becomes the policy's target in a distinct LAPA embodiment. The figure illustrates retrieval by similarity in that representation. e05e07e16
What it supports. These examples make the proposed bridge from human video to robot experience concrete: motion-related embeddings can group different embodiments without requiring matching joint vectors. This is qualitative evidence for shared semantics. The downstream augmentation results are a separate test of whether such supervision improves an action policy.
Where the evidence stops. The selected retrievals provide no accuracy measure, sampling distribution or guarantee of physical equivalence. The left/right labels describe the authors' examples, not a universal image-coordinate convention. A latent label also does not by itself specify executable commands for an arbitrary robot.
5.2 Turn future frames into offline supervision
The paper uses future images during data preparation in two different ways. The latent-action model learns a compact description of the transition between current and future frames by reconstructing the latter. Its continuous embedding becomes an action-like target for the LAPA embodiment. The robot-specific IDM instead predicts the actual intervening action chunk from a frame pair. Human videos provide latent targets; generated robot videos can receive either label type. A video generator expands the set of trajectories, and a multimodal judge filters or re-captions instruction mismatches. These auxiliary systems prepare training examples. At deployment, the policy conditions on current observations and state and generates actions through the DiT; the described control path does not require a future video or online inverse-dynamics plan. e03e04e05e07e09
5.3 Separate better averages from isolated mechanisms
Reader analysis: three questions need different evidence. Tables 2 and 3 ask whether the complete pretrained system adapts better than the selected baselines. Figure 9 asks whether neural trajectories help an already pretrained GR00T policy. Neither directly establishes that the dual-system split alone causes the gain. Appendix results also qualify the broad improvement narrative. At 100 demonstrations, DexMG Box Cleanup reports 29.4% for GR00T versus 80.8% for Diffusion Policy, despite the favorable suite average. GR-1's average changes from 50.0% at 100 demonstrations to 49.3% at 300, so growth is not uniformly monotonic. Together with the contradictory DexMG summary and missing uncertainty intervals, these details favor task-level, protocol-matched reproduction over treating a single overall score as a universal capability measure. e12e13e14e15e16e22
5.4 Training and inference
During training
Pre-training mixes real robot actions, latent actions from robot/human videos, and latent/IDM labels from generated videos. Table 7 reports 8,375.7 hours total: 3,288.8 robot, 2,517.0 human, 1,742.6 simulation and 827.3 neural-generated hours. These are source-reported corpus statistics, not equal sampling weights. e07e10
Both stages freeze the language component while tuning the vision encoder, DiT and adapters. Table 6 gives AdamW, learning rate 1e-4, cosine decay, 0.05 warmup, pre-training batch 16,384 for 200,000 steps, and post-training batches 128 or 1,024 for 20,000–60,000 steps. e07
Appendix F adds squared error on normalized target-object centers, labeled by OWL-v2, to the action loss. Post-training specializes to one embodiment; neural augmentation uses a 1:1 sampling ratio and restricts video-model/IDM training to available low-data demonstrations. e07e08e09
During inference
From current observations and state, the policy initializes Gaussian action noise and performs four denoising steps to produce a 16-action chunk. The 2.2B-parameter model reports 63.9 ms per chunk on an L40 using bf16. This is a model-sampling measurement; Section 1 separately describes 10 Hz System 2 and 120 Hz motor generation, without a complete deployment scheduling specification. e02e03e04
5.5 Implementation flow
- Encode the instruction and scene
Eagle-2 combines a SigLIP-2 image encoder with SmolLM2. Images at 224×224 become 64 tokens per frame after pixel shuffle. GR00T-N1-2B extracts the 12th language-model layer, which the authors report as faster and more effective than final-layer features. e03
- Condition motor generation
Embodiment-specific MLPs map state and noised actions into shared embeddings; the action encoder also receives diffusion time. DiT self-attention mixes state/action tokens, while cross-attention receives vision-language features. A corresponding MLP decodes the final action tokens. e03e04
- Construct video supervision
A VQ-VAE encodes current/future frames and reconstructs the future from the current frame plus a latent action. Policy training uses its continuous pre-quantization embedding as the distinct LAPA embodiment. A separate embodiment-specific IDM predicts intervening action chunks from two images using SigLIP-2 and a flow-matching DiT. e05e09
- Expand demonstrations
DexMimicGen transforms object-relative demonstration segments into new configurations, interpolates connections and retains successful replays. Neural trajectories instead come from a teleoperation-finetuned video generator with new prompts; filtering and re-captioning address instruction mismatch. These are offline training-data procedures. e06e09
6. Experiments & results
GR00T N1 couples Eagle-2 vision-language features to an embodiment-aware flow-matching action transformer. Human videos, simulated demonstrations and generated robot videos expand its training supervision. The strongest evidence is improved post-training performance on tabletop manipulation; the paper does not establish general humanoid locomotion or inference-time world-model planning.
6.1 Read the original evidence
Table 2. Post-training improves the reported benchmark averages, with a DexMG discrepancy in the appendix. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read down a benchmark column to compare methods under the stated 100-demonstration-per-task regime. RoboCasa contains 24 kitchen tasks, DexMG nine bimanual tasks, and GR-1 24 humanoid tabletop tasks. These columns therefore represent different task sets and embodiments. The last column is the paper's reported aggregate; do not assume that it is the arithmetic mean of the three preceding cells. The evaluation selects the maximum score among the final five checkpoints, each assessed over 100 trials. GR00T begins from heterogeneous pre-training, while the listed comparison policies are described as trained from scratch for these experiments. e11e12e13e15e22
What it supports. The GR-1 column reports 50.0% for GR00T versus 32.7% for Diffusion Policy, a 17.3-percentage-point difference. RoboCasa reports 32.1% versus 25.6%, a 6.5-point difference. These are simulated execution results under the paper's checkpoint-selection protocol, rather than a measurement of generated-video realism.
Where the evidence stops. DexMG reads 66.5% versus 56.1% here, but Table 4 and Figure 10 read 58.5% versus 46.9% at 100 demonstrations. The source gives no explanation. Preserve both versions and treat the aggregate as reported, pending reconciliation; no uncertainty intervals accompany this table.
Table 3. GR-1 robot evaluation shows an advantage at both post-training data budgets. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First compare methods at the same data budget: the first and third rows use 10%, while the second and fourth use the full task dataset. Then compare the two GR00T rows to see what additional task demonstrations provide. The budget refers to post-training demonstrations; GR00T retains its earlier pre-training. Columns group distinct physical behaviors, and Appendix Table 5 further separates seen and unseen objects for pick-and-place. The score allows partial progress. Most tasks use ten trials, but machinery packing counts objects placed within 30 seconds over five trials. Read the percentages using those scoring rules. e12e14
What it supports. GR00T's full-data score is 76.8%, compared with Diffusion Policy's 46.4%; the difference is 30.4 points. With 10% data, the corresponding scores are 42.6% and 10.2%. Low-data GR00T comes within 3.8 points of full-data Diffusion Policy, supporting the paper's adaptation-efficiency claim.
Where the evidence stops. The aggregate combines partial-credit tasks and an object-count task, so 76.8% cannot be read as the fraction of episodes completing every requirement. Small trial counts and absent uncertainty estimates limit precision. The comparison also does not isolate architecture from the benefit of pre-training.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| GR-1 real-world tabletop manipulation Post-training on full versus 10% task demonstrations; pick-and-place, articulated, industrial and coordination categories. Usually 10 trials per task with partial credit; machinery packing uses five trials and object counts within 30 seconds. | Full data: 76.8%; 10% data: 42.6%. Reported average policy score (%) under the paper's partial-credit protocol | Diffusion Policy: 46.4% full and 10.2% low-data; differences are 30.4 and 32.4 percentage points. Evidence for adaptation efficiency, not binary completion on every episode. Low-data GR00T remains 3.8 points below full-data Diffusion Policy. Baselines do not share GR00T's pre-training. e12e14 |
| Simulation manipulation with 100 demonstrations per task 24 RoboCasa, nine DexMG and 24 GR-1 tasks; 100 evaluation trials and the maximum over the final five checkpoints, saved 500 steps apart. | Table 2: RoboCasa 32.1, DexMG 66.5, GR-1 50.0; reported overall 45.0. Average success rate (%) | Diffusion Policy: 25.6, 56.1, 32.7; overall 33.4. BC Transformer: 26.3, 53.9, 16.1; overall 26.4. GR-1 improves by 17.3 points over Diffusion Policy. DexMG is internally inconsistent: Table 4 and Figure 10 instead report 58.5 versus 46.9 at 100 demonstrations. No reconciliation or revised aggregate is inferred. e11e12e13e15e22 |
| RoboCasa neural-trajectory post-training ablation 24 tasks; 30/100/300 demonstrations per task plus 3,000 generated trajectories per task; same pretrained GR00T starting point. | IDM labels: 21.2/40.9/56.4; LAPA labels: 21.6/39.5/52.9. Average success rate (%) | Without neural trajectories: 17.4/32.1/49.6. Both label types help here. LAPA leads slightly at 30 demonstrations; IDM leads with more demonstrations. This tests augmentation, without isolating each pre-training data source or proving pseudo-label accuracy. e07e12e16 |
| Real GR-1 neural-trajectory post-training ablation 10% demonstration regime; five seen-object pick-and-place and three industrial tasks, adding 100 generated trajectories per task. | 39.63 with IDM-labeled neural trajectories. Reported average partial-credit policy score (%) | 33.78 without augmentation; 4.07 for Diffusion Policy in this eight-task subset. The displayed Overall scores imply a 5.85-point gain; the text reports approximately 5.8. This subset excludes articulated and coordination tasks and is not Table 3's overall average. Its aggregation is also unresolved: weighting the displayed pick-and-place and industrial scores by the stated five and three tasks gives approximately 34.13 without augmentation and 40.00 with IDM, rather than the printed 33.78 and 39.63. These calculations are diagnostic, not replacement experimental results; the printed Overall values are retained without assuming an undocumented weighting rule. e12e17 |
| Pretrained GR-1 handover and novel-object placement No task-specific post-training; five objects and three trials per object for each of two settings. | Handover-to-shelf: 76.6% (11.5/15); novel object into unseen container: 73.3% (11/15). Partial-credit score (%) | No matched baseline is reported for these probes. A correct grasp followed by failed placement earns 0.5. These small targeted probes support limited transfer, not broad zero-shot reliability. e21 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 9. Neural trajectories add value, but the useful label representation depends on the data regime. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat the panels as separate experiments and reread each legend. On the left, green means LAPA labels and blue means IDM labels; on the right, green means IDM. RoboCasa varies the number of demonstrations per task and adds 3,000 generated trajectories per task. The robot panel fixes the 10% demonstration regime, adds 100 generated trajectories per task, and covers only five seen-object pick-and-place plus three industrial tasks. Dark gray is GR00T without this post-training augmentation. Comparing colored bars against dark gray therefore asks about additional neural supervision, whereas comparison against light-gray Diffusion Policy also changes the policy and pre-training. e07e12e16e17
What it supports. At 100 RoboCasa demonstrations, IDM augmentation raises success from 32.1% to 40.9%, an 8.8-point gain. LAPA is slightly stronger at 30 demonstrations, while IDM is stronger at 100 and 300. In the eight-task robot subset, the plotted average rises from 33.78% to 39.63%, a 5.85-point difference.
Where the evidence stops. These experiments do not isolate every pre-training data source and lack error bars or a matched-exposure control. The robot panel uses partial credit. Its printed Overall values are not recovered by weighting the two displayed categories by five and three tasks: the diagnostic GR00T/IDM means are about 34.13/40.00 rather than 33.78/39.63. The aggregation remains unresolved; the plotted values are retained as reported, not replaced by these calculations.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors limit current scope to short-horizon tabletop manipulation and identify physical consistency and diversity of synthetic trajectories as unresolved challenges. Section 4.5 also describes a handover skill lost after right-hand-only post-training. e18
The comparisons combine architecture, data and pre-training differences; they do not isolate the benefit of the dual-system design. Tables provide no uncertainty intervals, real-robot trial counts are small, and best-checkpoint simulation scoring differs from final-checkpoint evaluation. Table 4 also contains task-level reversals despite favorable benchmark averages. e03e12e13e14e15
The printed flow sign mismatch and conflicting DexMG summaries obstruct exact reproduction. Figure 3's frozen marker must be read with Section 2.3: the language component is frozen, while Table 6 explicitly unfreezes the vision encoder. e03e04e07e13e15e22
7.2 Questions for discussion
- Would neural augmentation still help with fixed real-data exposure and a matched resampling control? [e07, e16]
- How much handover generalization survives post-training when a small amount of pre-training data is replayed? [e18]
8. Reproducibility audit
8.1 Requirements and known gaps
The paper states that checkpoints, data and simulation benchmarks are released. Rebuilding pre-training would require the heterogeneous corpus, action-label generators and approximately 50,000 H100 GPU-hours. It reports training with up to 1,024 GPUs; single-A6000 finetuning was tested with adapter/DiT batch sizes up to 200, or up to 16 when tuning vision. e02e19
A faithful data pipeline needs the extended LeRobot modality specification, 6D state rotations, axis-angle action rotations, min-max scaling and left-to-right arm ordering. Appendix F identifies WAN2.1-I2V-14B with LoRA, but leaves the commercial filtering model/version, exact judgment prompts and detailed VQ-VAE configuration unspecified. Resolve the flow convention before implementation. e05e09e20e04
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: Check the flow direction before policy training
Reader-proposed check; not executed. Fix a ground-truth action chunk and a Gaussian noise sample, construct the paper's interpolation, and compare its finite-difference direction with the printed target. Then test the printed positive Euler update using that oracle target against an explicitly sign-consistent alternative, keeping four steps and the same endpoints. Measure endpoint distance to the action chunk. Increasing error under the printed pairing would expose the convention mismatch; a corrected toy trajectory would only establish mathematical consistency, not which convention the released implementation uses. e04
Check 2: Isolate what neural pseudo-labels contribute
Reader-proposed check; not executed. At RoboCasa's 30- and 100-demonstration regimes, start from the same GR00T checkpoint and compare no augmentation, resampled original demonstrations, and the same generated videos labeled by LAPA or IDM. Keep video count, evaluation configurations and initialization seeds fixed; add a control matching real-demonstration exposure because 1:1 mixing changes that exposure. Restrict generator and IDM fitting to the allowed demonstration subset. Report task-level success and uncertainty under both final-checkpoint and the paper's best-of-five rule. If augmentation loses its advantage against matched resampling, the evidence for new useful motion supervision weakens; a persistent IDM/LAPA crossover would support the proposed data-regime dependence. e07e09e11e12e16
8.3 Reading coverage
Visual audit: Visually inspected the title/byline, full main text and appendix pages 1–28, including Figures 1–14 and Tables 1–7. All five final original crops were inspected. Checked Figure 3 arrows, its frozen marker and denoising loop against Sections 2.1–2.3 and Table 6; checked Eq. (1) and the Euler update on page 5; checked Figure 9's panel-specific legends. Main/appendix DexMG discrepancies remain explicit. Figure 12's caption swaps the displayed object/row descriptions and names a destination inconsistent with the image labels; no task-specific conclusion is drawn from that caption [e23]. Reference pages 29–36 were read as text but not visually inspected. External videos, code and separate supplements were outside the supplied material.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28. 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
- 1. Introduction
- 2. GR00T N1 Foundation Model
- 2.1. Model Architecture
- 2.2. Training Data Generation
- 2.3. Training Details
- 3. Pre-Training Datasets
- 3.1. Real-World Datasets
- 3.2. Synthetic Datasets
- 3.3. Human Video Datasets
- 4. Evaluation
- 4.1. Simulation Benchmarks
- 4.2. Real-World Benchmarks
- 4.3. Experiment Setup
- 4.4. Quantitative Results
- 4.5. Qualitative Results
- 4.6. Limitations
- 5. Related Work
- 6. Conclusions
- A. Contributors and Acknowledgments (A.1–A.3)
- B. Detailed Experiment Results
- C. Additional Qualitative Results
- D. Hyperparameters
- E. System Design (E.1–E.2)
- F. Additional Training Details
- References
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.
- Identity/version: the title matches exactly. The supplied PDF identifies arXiv:2503.14734v2, 27 March 2025, and separately prints 2025-3-28 on its title page. The catalog submission date is 18 March 2025; this review concerns the supplied v2, without comparison to v1.
- Author metadata: the title-page author is NVIDIA, with a footnote directing readers to Appendix A. The catalog expands the collective credit into contributors and contains a standalone colon. Appendix A substantially matches those names, but prints Xinye (Dennis) Da where the catalog says Xingye Da; that spelling difference is unresolved. Metadata preserves the verified collective byline rather than asserting an individual-author ordering or unverified affiliations.
- All ten supplied text chunks were read, including references on pages 29–36. Pages 1–28 were also rendered and visually inspected, including all original figures and tables. References were read as text; cited external works were not independently inspected.
- The acquisition warning that text extraction does not reconstruct images was addressed by inspecting the retained PDF and every final crop. Separate supplemental material was not supplied and its availability remains unverified.
- Code, model checkpoints, external datasets and linked project resources were not inspected; no paper scripts or experiments were run.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/byline/date/arXiv margin and footnote; p. 19, Appendix A.1–A.2
Exact title and NVIDIA collective byline are printed. The margin identifies 2503.14734v2, 27 March 2025; the header reads 2025-3-28. Contributor roles include Xinye (Dennis) Da.
Go to primary source ↓e02PDF pp. 1–2, Introduction, Figure 1 and final paragraph
Describes fragmented robot data, the data pyramid, dual-system VLA, 10 Hz VLM and 120 Hz motor generation, and states that checkpoints, datasets and simulation benchmarks are publicly released.
Go to primary source ↓e03PDF pp. 3–4, Section 2/2.1, Figures 2–3 and System 2 subsection
2.2B parameters, 1.34B VLM, 63.9 ms per 16-action chunk on L40 bf16; Eagle-2/SigLIP-2/SmolLM2, 224×224 images, 64 image tokens, 12th-layer features, embodiment MLPs, cross-attention and the frozen marker.
Go to primary source ↓e04PDF p. 5, Section 2.1, Eq. (1), preceding interpolation and following Euler update
Defines state/action self-attention and vision-language cross-attention. Interpolation is tau A+(1−tau)epsilon; printed loss target is epsilon−A, while Euler advances tau by adding V/K. Uses four denoising steps.
Go to primary source ↓e05PDF p. 5, Section 2.2 Latent Actions; p. 6, Figure 4/caption
VQ-VAE learns from current/future frame pairs. Its continuous pre-quantization embeddings become flow-matching targets for a separate LAPA embodiment. Figure 4 retrieves examples labeled right-arm movement left/right across eight embodiments.
Go to primary source ↓e06PDF pp. 6–7, Section 2.2 Neural/Simulation Trajectories and Figure 5; p. 10, Section 3.2
Generated videos vary prompts from starting frames. DexMimicGen segments, transforms and interpolates object-centric demonstrations and retains successful replays; Section 3.2 describes randomized tabletop configurations.
Go to primary source ↓e07PDF p. 8, Section 2.3 Pre-training/Post-training; p. 27, Table 6
Specifies target types, a frozen language component, embodiment-specific post-training, limited-data restrictions and 1:1 real/neural sampling. Table 6 lists optimization settings and explicitly unfrozen vision encoder and DiT; its frozen row is labeled text tokenizer.
Go to primary source ↓e08PDF p. 24, Appendix F, Auxiliary Object Detection Loss
OWL-v2 labels the target-object bounding-box center. A linear head predicts normalized 2D coordinates with squared error, added to flow-matching loss.
Go to primary source ↓e09PDF p. 24, Appendix F, Neural Trajectory Generation and IDM Model Training
WAN2.1-I2V-14B uses LoRA and 81-frame 480P inputs; a commercial multimodal LLM judges eight sampled frames, and failures are re-captioned using 16 frames at 256P. A two-image SigLIP-2/DiT IDM trains per embodiment for 30K or 60K. Exact commercial model/version and judgment prompts are absent.
Go to primary source ↓e10PDF p. 28, Table 7, category totals and Total row
Reports 592.9M frames and 8,375.7 hours: robot 3,288.8, human 2,517.0, simulation 1,742.6 and neural-generated 827.3 hours. Counts do not specify sampling weights.
Go to primary source ↓e11PDF p. 12, Section 4.1, all three benchmark bullets
RoboCasa has 24 kitchen tasks, DexMG nine tasks across three bimanual embodiments, and GR-1 24 tasks including 18 rearrangements with unseen receptacle combinations and six articulated-object tasks.
Go to primary source ↓e12PDF p. 14, Section 4.3 Baselines/Evaluation Protocol; p. 15, opening continuation
Simulation uses 100 trials and the maximum of the last five checkpoints saved every 500 steps. Real-world evaluation uses partial scoring over ten trials, except five machinery-packing trials scoring five objects within 30 seconds; low-data subsets use 10%.
Go to primary source ↓e13PDF p. 15, Table 2, all method rows and benchmark/Average columns
BC: 26.3/53.9/16.1/26.4; DP: 25.6/56.1/32.7/33.4; GR00T: 32.1/66.5/50.0/45.0 percent for RoboCasa/DexMG/GR-1/Average, at 100 demonstrations per task.
Go to primary source ↓e14PDF p. 15, Table 3; p. 27, Table 5, average and seen/unseen pick-and-place rows
Overall DP 10%/full scores are 10.2/46.4; GR00T 42.6/76.8. Full-data GR00T categories are 82.0/70.9/70.0/82.5. Table 5 splits seen/unseen pick-and-place, with full-data GR00T 92.0/72.0.
Go to primary source ↓e15PDF p. 26, Table 4, DexMG Average and Box Cleanup rows; GR-1 Average row
At 100 demos DexMG averages are DP 46.9 and GR00T 58.5, conflicting with Table 2. Box Cleanup reverses the overall comparison (80.8 versus 29.4). GR-1 GR00T averages are 43.2/50.0/49.3 at 30/100/300 demos.
Go to primary source ↓e16PDF p. 16, Figure 9 left panel, caption and Section 4.4 continuation
RoboCasa no-augmentation GR00T: 17.4/32.1/49.6; LAPA: 21.6/39.5/52.9; IDM: 21.2/40.9/56.4 at 30/100/300 demos. Adds 3,000 generated trajectories per task; no uncertainty bars are shown.
Go to primary source ↓e17PDF p. 16, Figure 9 right panel, caption and accompanying results
With 10% demonstrations and 100 generated trajectories per task, Figure 9 prints Overall (8 Tasks) scores of DP 4.07, GR00T 33.78 and GR00T+IDM 39.63. Its five seen-object pick-and-place scores are 2.0/36.0/42.0, and its three industrial scores are 6.67/31.0/36.67, respectively. Reader analysis: a 5:3 task-weighted recomputation yields approximately 3.75/34.13/40.00, which does not reproduce the printed Overall values; the source does not explain the aggregation. The accompanying text describes a 5.8-point gain. Author values are preserved.
Go to primary source ↓e18PDF p. 16, Section 4.5 handover example; p. 17, Section 4.6
Reports loss of handover behavior after right-hand-only post-training. Authors restrict present scope to short-horizon tabletop tasks and discuss limitations in physically consistent, diverse synthetic generation.
Go to primary source ↓e19PDF p. 8, Section 2.3 Training Infrastructure
H100 cluster with up to 1,024 GPUs; about 50,000 H100 GPU-hours for GR00T-N1-2B pre-training. Single-A6000 finetuning tested with batch up to 200 for adapters/DiT or 16 when tuning vision.
Go to primary source ↓e20PDF pp. 23–24, Appendix E.1 Dataset Formats and E.2 Standardized Action Spaces
Extended LeRobot format declares modality semantics and rotation representations. Best-effort normalization uses 6D end-effector rotation states, axis-angle actions, min-max scaling and a left-to-right arm ordering.
Go to primary source ↓e21PDF p. 15, Section 4.4 Pre-training Evaluations
Two pretrained GR-1 probes use five objects, three trials each. Reports 76.6% (11.5/15) for handover-to-shelf and 73.3% (11/15) for novel object/container; correct grasp without placement earns half credit.
Go to primary source ↓e22PDF p. 20, Appendix B, Figure 10, DexMG middle group and GR-1 right group
Figure 10 agrees with Table 4's 100-demo DexMG averages, 46.9 for DP and 58.5 for GR00T, and shows GR-1 GR00T 43.2/50.0/49.3 across data regimes.
Go to primary source ↓e23PDF p. 21, Figure 12, task labels and caption
The upper example is labeled lemon from cutting board to basket and the lower cucumber from placemat to basket. The caption instead describes cucumber in the top example and lemon in the bottom, and names a pan for the latter.
Go to primary source ↓8.5 Primary sources
GR00T N1: An Open Foundation Model for Generalist Humanoid Robots ↗
PDF · 17,099 extracted words
Source fingerprint
76e38cf679637b816b16edf7170340883dd09175bf334aeeaed49392777731b9