FLARE: Robot Learning with Implicit World Modeling
1. Paper overview
In one sentence: Training a shared action transformer to anticipate compact future observations improves manipulation, while leaving explicit planning and the reliability of its latent dynamics untested. e01e02e04e05e07e10e11e20
| At a glance | What to know |
|---|---|
| Research problem | Author claim Can future prediction improve manipulation without expensive visual reconstruction? The authors propose supervising compact future features inside the policy, preserving action generation while avoiding a separate video-generation objective. e02 |
| Core mechanism | Source description Future tokens and action tokens interact within one DiT, allowing future alignment to shape the action policy through shared self-attention. e04 |
| A key reported result | 24 RoboCasa simulation tasks: 70.1% Average task success rate. In-domain target pretraining; 50 episodes per task every 1,000 updates; maximum over final five checkpoints. Policy-only 61.9%; UWM 60.8%; GR00T N1 (Scratch) 60.6%; Diffusion Policy 51.7%. 8.2 percentage points above policy-only, calculated from Table 1. UWM receives 400k updates versus 80k policy updates for others; FLARE also has target pretraining. e07e08 |
| Reading caution | Source description The authors restrict real-world testing mainly to imitation-based pick-and-place and controlled GoPro collection. Fine dexterity, reinforcement learning and naturalistic video transfer remain future work; novel objects still require expert robot demonstrations. e15 |
Core contributions
- Source description
Future tokens and action tokens interact within one DiT, allowing future alignment to shape the action policy through shared self-attention. e04
Figure 2. Future supervision enters the same transformer that denoises robot actions. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read upward from the bottom token sequence: proprioception, noised actions and learnable future tokens enter shared DiT self-attention. The left branch supplies the current image and instruction through cross-attention. Action loss is attached at the top, while the dashed alignment connection taps the future-token stream at an internal layer. The right-hand image is a recorded future observation supplying a training target, not an image available to the deployed robot. The snowflake marks a target without backpropagated gradients. Section 3.2 and the EMA update clarify that this encoder can still change between steps; Algorithm 1 confirms the no-gradient target computation. e03e04e06e18e20
What it supports. The architecture provides a route for future supervision to affect action generation through shared attention. It does not require the robot to render future frames or run an external inverse-dynamics model. This supports describing FLARE as an implicit future-prediction objective inside a policy, rather than treating the drawing as an explicit rollout planner.
Where the evidence stops. Figure 2 labels actions through t+H−1, whereas Section 2 includes t+H. Equation (1) and Algorithm 1 also disagree on velocity sign. These unresolved source differences affect implementation; the crop is preserved unchanged.
2. Motivation
2.1 The problem and the proposed response
Can future prediction improve manipulation without expensive visual reconstruction? The authors propose supervising compact future features inside the policy, preserving action generation while avoiding a separate video-generation objective. e02
2.2 What this reading follows
A robot policy can learn from what happens next without rendering the next image. FLARE adds future tokens to an action-denoising transformer and trains them to match embeddings of later observations. A compact encoder first learns which visual and language information helps predict actions; its features then become the policy's future targets. This illustrated reading separates the architecture from three kinds of evidence: in-domain simulation comparisons, posttraining on limited robot data, and transfer from action-free human videos. The reported improvements are substantial, but checkpoint selection, partial-credit scoring and inconsistencies within this v1 PDF matter when interpreting or reproducing them. e01e02e04e05e07e10e11e20
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 | Not assigned |
| Architecture | Not assigned |
| Prediction paradigm | Not assigned |
| Quadrant | Not assigned |
This table preserves the labels recorded at reading time. The current major category is WAMs. View the current classification.
3.1 Evidence-based assessment
Insufficient evidence to decide
The catalog has no assigned classification to support or contradict. Architecture evidence supports one shared policy DiT with action and future-token streams plus a training target encoder. This is future-latent alignment coupled to action prediction, without inverse dynamics or demonstrated inference-time world-model planning. e04e20
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 First learn what an observation must preserve for action
FLARE begins by training the representation that will later define a useful prediction target. SigLIP2 provides visual and text features, but the proposed encoder does not simply retain all patch tokens. Four transformer fusion layers combine the modalities, and a Q-former produces 32 learned query outputs. During this stage, an attached eight-block DiT predicts demonstrated actions using those outputs. This makes action prediction the pressure shaping the compressed representation. The distinction between experiments is essential: the simulation comparison first learns this encoder only from the same benchmark's data, whereas the transfer experiments start from the cross-embodiment mixture. Table 2 then asks whether generic visual targets can substitute for the proposed encoder; they help, but their reported scores are lower. e05e07e09e12
5.2 Use future observations to supervise shared policy computation
The next stage inserts future tokens beside the policy's state and action tokens. These streams share self-attention, and current visual-language features enter through cross-attention. A projected internal-layer representation predicts the target features of a later observation, while the action stream still learns denoising. Reader interpretation: this supplies a representation-level incentive to preserve consequences relevant to control, but does not prove the tokens encode a correct causal simulator. The future image belongs to training supervision; inference still starts from noisy actions and uses four refinement steps. The target branch is stopped for gradients and can follow EMA updates. Implementation needs care: the stated velocity target and pseudocode disagree in sign, and the pseudocode omits the explicit intermediate-layer selection and EMA update. e03e04e06e18e20
5.3 Trace each gain to the experiment that actually measures it
The simulation table tests control performance on two multitask suites after in-domain training. The posttraining figure instead tests transfer from a cross-embodiment encoder as robot data varies, and separately reports controlled physical execution. Finally, the human-video experiment adds alignment-only examples and asks whether a few robot demonstrations suffice for new object geometries. Reader interpretation: these experiments collectively support the usefulness of the training signal more directly than the stronger claim of general world understanding. The human-video bars also include half credit for successful grasping without placement, so they cannot be read as full completion probabilities. The static appendix rollouts illustrate the reported behaviors but do not independently establish continuous trajectories, latency or robustness beyond the tested initial conditions. e07e08e09e10e11e19
5.4 Training and inference
During training
Section 4.1 learns target embeddings only from each benchmark's own data for 80,000 steps. Section 4.2 instead uses cross-embodiment pretraining: FLARE warm-starts its encoder, whereas policy-only also initializes its DiT. These are distinct comparison regimes. e07e09
Robot examples use action flow matching plus alignment weighted by 0.2. Human GoPro examples use alignment alone, mixed with limited robot demonstrations and the GR1 pretraining dataset; the paper does not specify their missing action/proprioception input handling. e04e11e18
During inference
Sample a Gaussian action chunk and refine it with four Euler denoising steps conditioned on current observation and proprioception. The described policy directly outputs actions; it does not search candidate latent rollouts. Executed-chunk length and observation refresh timing are not fully specified. e03e20
5.5 Implementation flow
- Compress observations for control
SigLIP2-large-patch16-256 encodes images and instructions. For the single-image description, 256 image tokens and 32 text tokens pass through four fusion layers, then a Q-former compresses them to 32 queries. Attaching eight DiT blocks and training action prediction makes the representation action-aware. e05
- Mix action and future tokens
Concatenate encoded proprioception, noisy actions and 32 learnable future tokens. DiT self-attention mixes these streams; cross-attention supplies current vision-language context. Project future-token activations from layer 6 of 8 through an MLP and align them with the future observation target. e04e05e13
- Separate target supervision from adaptation
The future encoder receives no alignment gradient. During cross-embodiment posttraining it follows the policy encoder by EMA, with retention coefficient 0.995. Thus Figure 2's snowflake means a gradient-stopped target, not necessarily permanently fixed weights. e04e06e14e18
6. Experiments & results
FLARE trains a flow-matching robot policy to match future observation embeddings inside its action-denoising transformer. Compact, action-trained visual-language targets supply an auxiliary learning signal, including from action-free human videos. Simulation and real-robot improvements support this training recipe; they do not establish an explicit planner or calibrated world simulator.
6.1 Read the original evidence
Table 1. In-domain future alignment improves the printed aggregate scores on both simulation suites. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the two bold benchmark-average rows; they describe separate robot embodiments and should not be pooled into one score. Compare FLARE with Policy Only to examine the benefit relative to the authors' own policy architecture, then inspect UWM and GR00T N1 (Scratch) for broader comparisons. The surrounding protocol matters: these targets were pretrained on in-domain data, not the cross-embodiment mixture. Each policy checkpoint receives 50 episodes per task, and the maximum success rate over the final five checkpoints is reported. Most policies train for 80,000 updates; UWM trains for 400,000, and FLARE additionally requires target-encoder pretraining. e07e08
What it supports. FLARE reaches 70.1% versus 61.9% policy-only on RoboCasa, and 55.0% versus 44.0% on GR1. The corresponding gains are 8.2 and 11.0 percentage points, calculated from the table. These are executed simulation-task outcomes, providing stronger control evidence than visual prediction quality alone.
Where the evidence stops. The table supplies no seed-level uncertainty. GR1 category rows do not transparently recover all aggregate rows under the stated task counts; retain printed aggregates without inventing a correction. The comparison is not matched for total training compute.
Figure 6. Cross-embodiment target pretraining transfers to simulation and controlled real-robot tasks. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the panels independently. On the left, each x-axis group specifies trajectories per task across 24 RoboCasa tasks; the bars compare the same data budget. On the right, the first pair is an average, followed by the four source-to-destination real-robot tasks, not additional data-budget settings. The policy-only baseline initializes both encoder and DiT from pretraining, whereas FLARE initializes only its encoder. The real experiments use 100 robot trajectories per task and the final checkpoint, with eight reference initial frames per task involving four object types. RoboCasa instead uses the checkpoint-selection protocol from Section 4.1. e08e09e10e15
What it supports. RoboCasa rises from 42.3% to 52.1% at 100 trajectories per task and from 65.3% to 71.3% at 1,000. The real-robot average is labeled 81.2% versus 95.3%. The placemat-to-basket pair is tied at 87.5%, showing that the average gain does not imply improvement on every task.
Where the evidence stops. Section 4.2 prose says 95.1%, conflicting with the plotted 95.3%; its in-domain RoboCasa reference says 70.2%, versus Table 1's 70.1%. No uncertainty bars or reconciliation are supplied, and these are controlled pick-and-place evaluations.
Figure 7. Action-free human videos improve a partially credited novel-object manipulation score. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the train/test diagram: the novel objects receive human demonstrations and a few robot demonstrations before real-robot testing. They are novel relative to the earlier dataset, not objects withheld from adaptation. The two right-hand groups count robot trajectories per object; they do not count the additional human videos. Section 4.3 specifies 150 GoPro demonstrations for each of five objects, plus GR1 pretraining data. Robot examples train both losses, while human examples train only future alignment. The y-axis reads 'Average Performance'; the chart title reads 'Success Rate on Novel Objects'. The protocol awards half credit when grasping succeeds but basket placement fails, so interpret the bars as scores with partial credit. e11e15e18
What it supports. With one robot trajectory per object, adding human videos raises the reported score from 37.5% to 60.0%; with ten it rises from 42.5% to 80.0%. These comparisons support useful transfer from action-free video under the stated adaptation protocol, while still requiring robot demonstrations of each object.
Where the evidence stops. The paper describes five reference poses per object but does not fully specify repeat counts or human-example action/proprioception inputs. Scores include partial grasps, and controlled GoPro collection does not establish transfer from arbitrary internet video.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| 24 RoboCasa simulation tasks In-domain target pretraining; 50 episodes per task every 1,000 updates; maximum over final five checkpoints. | 70.1% Average task success rate | Policy-only 61.9%; UWM 60.8%; GR00T N1 (Scratch) 60.6%; Diffusion Policy 51.7%. 8.2 percentage points above policy-only, calculated from Table 1. UWM receives 400k updates versus 80k policy updates for others; FLARE also has target pretraining. e07e08 |
| 24 GR1 simulation tasks 18 rearrangement and 6 articulated tasks; same checkpoint-selection protocol as RoboCasa. | 55.0% Average task success rate | Policy-only 44.0%; UWM 29.5%; GR00T N1 (Scratch) 45.1%; Diffusion Policy 40.9%. 11.0 percentage points above policy-only. These are printed aggregate rows; category rows do not transparently reconstruct all aggregates, so no alternative averaging is invented. e07e08 |
| RoboCasa posttraining with a cross-embodiment encoder 24 tasks with 100, 300 or 1,000 trajectories each; Section 4.1 evaluation protocol. | 52.1%, 66.4%, 71.3%, respectively. Average task success rate | Policy-only: 42.3%, 60.1%, 65.3%. Gains persist across these data budgets. The p. 7 comparison quotes an in-domain result of 70.2%, whereas Table 1 prints 70.1%; preserve both locations. e09e10e08 |
| Four real GR1 pick-and-place tasks 100 trajectories per task; eight reference initial frames per task involving apple, can, bottled water and cucumber; final checkpoint. | 95.3% in Figure 6; 95.1% in Section 4.2 prose. Average policy success rate | Figure 6 policy-only average: 81.2%. Physical execution evidence under a controlled setup. The conflicting printed averages remain unresolved, and the compared DiT initializations differ. e09e10 |
| Novel-object transfer from human egocentric videos Five objects; 150 human videos per object, GR1 pretraining data, and 1 or 10 robot demonstrations per object; five reference initial poses per object. | 60.0% with 1 robot demonstration; 80.0% with 10. Reported success score, with 0.5 credit for grasping without successful basket placement | Without human videos: 37.5% and 42.5%, respectively. 22.5 and 37.5 percentage-point gains, calculated. These scores include partial credit and are not binary full-task completion rates or zero-shot transfer. e11 |
| GR1 target-embedding ablation Section 4.4 compares no alignment, generic SigLIP2 targets and action-aware targets. | 55.0% with action-aware embeddings. Success rate | No FLARE loss 43.9%; SigLIP2 49.6%; average-pooled SigLIP2 50.9%. Target choice matters, but encoder training and token count both change; this does not isolate action-awareness alone. e12 |
| RoboCasa EMA retention ablation 24 tasks × 300 trajectories. | 66.4% at rho=0.995. Policy success rate | Policy-only 60.1%; rho=0.99: 63.5%; 0.999: 65.7%; fixed target (1.0): 64.6%. Alignment helps even with fixed targets; the best observed EMA setting adds 1.8 percentage points over that fixed-target result. e14 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 2. The future target's representation changes the size of the policy improvement. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read downward from No FLARE loss to establish the ablation's own baseline of 43.9%. The generic SigLIP2 rows supervise future visual features at t+16: the unpooled variant uses 256 tokens per image, while 2×2 average pooling reduces this to 64. The final row switches to the compact representation trained for action prediction, described in Section 3.2 and Appendix A. It uses 32 learned queries after visual-language fusion. This sequence tests whether target design matters to FLARE, but it changes both representation training and token organization, so it cannot assign the entire improvement to a single architectural ingredient. e05e08e12
What it supports. Both generic visual targets improve on the no-alignment result, so FLARE's benefit is not confined to the proposed action-aware encoder. The best listed target reaches 55.0%, compared with 50.9% for pooled SigLIP2. That 4.1 percentage-point difference motivates a controlled target-encoder study.
Where the evidence stops. Do not silently replace this table's 43.9% baseline with Table 1's 44.0%. No variability is reported. Teacher training, compactness and target semantics vary together, limiting a causal interpretation of action-awareness alone.
Figure 9. A slowly adapting future target performs best in this RoboCasa sweep. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The horizontal labels denote retention coefficients for the target encoder, except Baseline, which removes future alignment altogether. Section 4.4 defines each update as retaining rho times the old target parameters and adding one minus rho times the policy encoder parameters. Therefore 1.0 freezes the target, while 0.99 changes it more quickly than 0.995 or 0.999. All comparisons here use 300 trajectories for each of 24 RoboCasa tasks. Read the numerical labels rather than bar-height ratios: the vertical axis starts at 50%, which visually magnifies the differences but does not change their reported values. e06e14
What it supports. The best plotted value, 66.4% at 0.995, exceeds the fixed-target result of 64.6% by 1.8 percentage points. Every alignment variant also exceeds the 60.1% policy-only baseline. Thus adaptation helps in this sweep, while future alignment remains useful even without adaptation.
Where the evidence stops. The authors attribute the poorer 0.99 result to instability, but this plot measures success, not target instability directly. Without repeated-seed uncertainty, small differences between retention settings should not be treated as universally optimal.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors restrict real-world testing mainly to imitation-based pick-and-place and controlled GoPro collection. Fine dexterity, reinforcement learning and naturalistic video transfer remain future work; novel objects still require expert robot demonstrations. e15
Reported tables and bars lack seed-level uncertainty. Best-of-final-five simulation selection differs from final-checkpoint real evaluation. Static rollout panels support qualitative inspection, not measured latency, full-video verification or proof of explicit future reasoning. e07e09e19
7.2 Questions for discussion
- Would temporally shuffled future targets retain the control improvement?
- How much human-video benefit survives evaluation using binary placement success on independently reset poses?
8. Reproducibility audit
8.1 Requirements and known gaps
Appendix C specifies pretraining on 256 H100 GPUs, batch 8,192, 150k steps; multitask training on 32 H100s, batch 1,024, 80k steps. AdamW uses betas (0.95, 0.999), epsilon 1e-8, weight decay 1e-5 and cosine scheduling with 0.05 warmup. Peak learning rate is absent. e17
Reproduction needs the GR1 in-house/simulation and seven OXE datasets. Section 3.2 says approximately 2,000 hours, but Table 3 totals 169.5M frames and 2,989.5 hours. Resolve this discrepancy rather than choosing an undocumented total. e16
Algorithm 1 omits explicit intermediate-layer extraction and EMA updates. Figure 2 labels actions through t+H−1, while Section 2 defines a chunk through t+H; the SigLIP2 ablation alone specifies t+16. Resolve indexing, velocity sign, human-input masking and sampling ratios before claiming faithful reproduction. e03e04e11e12e18
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 the correct future matter beyond extra supervision?
Proposed, not run: on the 24 GR1 simulation tasks, compare policy-only, correct future targets, current-observation targets and within-task shuffled future targets. Hold the action-aware encoder, token count, model initialization, data, alignment layer, loss weight and total training budget fixed. Use repeated seeds and identical evaluation episodes; report both final-checkpoint and source-style selected-checkpoint results. A reproducible advantage for correctly paired future targets over both controls would support temporal prediction as the mechanism. Similar gains from current or shuffled targets would weaken that explanation while preserving an auxiliary-supervision benefit. e04e05e07e12e13
Check 2: Does ordered human video improve full placement?
Proposed, not run: repeat the five-object adaptation experiment at one and ten robot demonstrations, comparing no human data, ordered GoPro trajectories and the same human frames with future pairing shuffled within each object. Keep robot examples and optimization budgets matched, and explicitly document the absent-action/proprioception input convention. Evaluate repeated, independently reset poses, reporting grasp success, binary placement success and the paper's partial-credit score separately. Improvement only in partial grasp credit would narrow the transfer claim; an ordered-video advantage in binary placement would support the proposed temporal learning benefit. e11e15e18
8.3 Reading coverage
Visual audit: Visually inspected the title/author/version page; every main-text method, evaluation, result and limitation page; Appendix A architecture, B data table, C hardware/training settings, D pseudocode, and E static rollout panels. Figures 1–12 and Tables 1–3 were inspected in their original PDF pages. All six final crops were separately viewed with labels and legends intact. Figure 7's 'Average Performance' y-axis and 'Success Rate on Novel Objects' title were checked against the partial-credit protocol on page 7. Architecture arrows and the target snowflake were cross-checked against Sections 3.1–3.2, the EMA update and Algorithm 1; unresolved sign, horizon-indexing and schematic omissions are disclosed. Numerical discrepancies between Table 1, Figure 6 and Section 4.2, and the main-text/appendix data totals, are preserved. Pages 15–20 contain references and were read in the complete text chunks, without a separate image pass. No linked videos, code or separate supplements were inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 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, author block and abstract (p. 1)
- Sections 1–3: introduction, flow matching and FLARE method (pp. 1–4)
- Sections 4.1–4.4: simulation, real-robot transfer, human videos and ablations (pp. 4–9)
- Sections 5–7 and acknowledgement (pp. 9–10)
- Appendices A–D: embedding architecture, data mixture, training and pseudocode (pp. 11–12)
- Appendices E.1–E.2: real-robot rollout panels (pp. 13–14)
- References (pp. 15–20)
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 complete supplied 20-page PDF text was read in all five chunks. Identity matches the catalog: arXiv:2505.15659v1 [cs.RO], 21 May 2025; title and all 21 authors agree. No other revision or edition was supplied or compared.
- Text extraction does not reconstruct figure images; this omission was addressed by inspecting original PDF pages 1–14 and all six final crops. Reference pages 15–20 were read as text.
- Separate supplemental material availability has not been fully verified; no separate supplement was supplied. Linked project media and code were not inspected, and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block and arXiv margin
Exact title and all 21 catalog authors match. The artifact is arXiv:2505.15659v1 [cs.RO], 21 May 2025. Affiliations are NVIDIA; University of Maryland, College Park; Nanyang Technological University; University of Texas, Austin.
Go to primary source ↓e02PDF pp. 1–2, Section 1 and Figure 1
The authors motivate compact future-observation alignment as a way to avoid visual reconstruction while improving policy training and accepting action-free video.
Go to primary source ↓e03PDF pp. 2–3, Section 2, Eq. (1) and Euler update
Defines observations, proprioception, expert action chunk through t+H, interpolation from noise to actions and four Euler steps. The printed velocity target is epsilon minus actions.
Go to primary source ↓e04PDF pp. 3–4, Figure 2, Section 3.1, Eqs. (2)–(3)
State, action and future tokens share DiT self-attention; current observations provide cross-attention. Projected intermediate future tokens match stopped-gradient future targets via cosine alignment weighted by 0.2. Figure labels the action range through t+H−1.
Go to primary source ↓e05PDF p. 4, Section 3.2; p. 11, Appendix A and Figure 10
Action flow-matching pretraining uses an eight-block DiT. SigLIP2 image/text features undergo four fusion layers and Q-former compression to 32 queries; Appendix A supplies backbone, resolution and single-image token counts.
Go to primary source ↓e06PDF p. 4, Section 3.2 and Figure 3; p. 9, Section 4.4 EMA update
Cross-embodiment pretrained embeddings initialize the policy encoder and provide future targets, adapted by EMA from the policy encoder with preferred retention 0.995.
Go to primary source ↓e07PDF pp. 4–5, Section 4.1 and Figure 4; p. 6, Section 4.1 continuation
RoboCasa uses three camera views; GR1 uses a head camera and 18 rearrangement plus 6 articulated tasks. In-domain target pretraining lasts 80k steps; policy training is 80k, UWM 400k. Evaluation uses 50 episodes per task every 1k updates and the maximum of the final five checkpoints. The extended policy-only run reports 44.1% at 160k.
Go to primary source ↓e08PDF p. 5, Table 1, all method columns and benchmark-average rows
RoboCasa averages are 70.1/61.9/60.8/60.6/51.7%; GR1 averages are 55.0/44.0/29.5/45.1/40.9%, ordered FLARE, policy-only, UWM, GR00T N1 (Scratch), Diffusion Policy. Category rows are separately printed; their aggregation is not clarified.
Go to primary source ↓e09PDF p. 2, Section 1 real-robot data statement; pp. 6–7, Section 4.2 and Figure 5
Real-robot posttraining uses 100 trajectories per task. FLARE initializes only its encoder; policy-only also initializes its DiT. Real evaluation uses eight reference initial frames per task involving four named objects and the final checkpoint; RoboCasa follows Section 4.1.
Go to primary source ↓e10PDF p. 6, Figure 6, both panels; p. 7, Section 4.2 continuation
RoboCasa FLARE versus policy-only is 52.1/42.3, 66.4/60.1 and 71.3/65.3% at 100/300/1000 trajectories per task. Real averages are labeled 95.3/81.2%; real FLARE task bars are 100.0, 93.8, 100.0, 87.5%. Prose instead says 95.1% and cites an in-domain RoboCasa comparison of 70.2%.
Go to primary source ↓e11PDF p. 7, Figure 7 and Section 4.3
Five novel objects, 150 GoPro demonstrations per object, GR1 pretraining data and limited robot examples form the transfer experiment. Human video uses only alignment. Figure 7's y-axis reads 'Average Performance' and its title reads 'Success Rate on Novel Objects'. Five initial poses per object are described; grasp without basket placement earns 0.5 credit. With/without human-video scores are 60.0/37.5% at one robot example and 80.0/42.5% at ten.
Go to primary source ↓e12PDF p. 8, Section 4.4 target-encoder paragraph and Table 2
No alignment/SigLIP2/pooled SigLIP2/action-aware scores are 43.9/49.6/50.9/55.0%. Generic targets use t+16, with 256 raw or 64 pooled tokens per image; action-aware targets use a different, policy-trained representation.
Go to primary source ↓e13PDF p. 8, Figure 8 and layer/coefficient paragraph
The GR1 sweep supports the selected layer 6 of 8 and alignment coefficient 0.2; earlier-layer alignment lowers reported success. The two panels respectively vary layer and coefficient despite a duplicated 'Left' caption label.
Go to primary source ↓e14PDF p. 8, Figure 9 and caption; p. 9, Section 4.4 EMA paragraph and update
For 24×300 RoboCasa trajectories, baseline/0.99/0.995/0.999/1.0 scores are 60.1/63.5/66.4/65.7/64.6%. Rho retains previous target parameters; 1.0 fixes the target. Instability from frequent updates is the authors' explanation, not a separately measured quantity.
Go to primary source ↓e15PDF p. 10, Section 6
Authors identify pick-and-place imitation, continued need for expert demonstrations and controlled head-mounted GoPro data as limitations; dexterity, reinforcement learning and natural video remain future directions.
Go to primary source ↓e16PDF p. 4, Section 3.2 data-total statement and Figure 3; p. 11, Appendix B, Table 3
Main text states approximately 2,000 hours. Appendix lists GR1 in-house, GR1 simulation and seven OXE datasets, with a total of 169.5M frames and 2,989.5 hours. The discrepancy is not explained.
Go to primary source ↓e17PDF pp. 11–12, Appendix C
Reports 256 H100s, batch 8192, 150k pretraining updates; 32 H100s, batch 1024, 80k multitask updates. Gives AdamW betas, epsilon, weight decay, cosine scheduling and warmup ratio, but no peak learning rate.
Go to primary source ↓e18PDF p. 12, Appendix D, Algorithm 1; p. 3, Eq. (1) and Section 3.1; p. 9, EMA update
Pseudocode uses actions minus noise, a no-gradient future target and one-minus-cosine loss. It does not explicitly implement the stated intermediate-layer tap, EMA update or action-free data branch. Its velocity sign differs from Eq. (1).
Go to primary source ↓e19PDF pp. 13–14, Appendices E.1–E.2, Figures 11–12
Static sequences illustrate four pick-and-place tasks and five novel objects. The text describes maneuvering around bottles/cans; these panels provide no timing or statistical uncertainty.
Go to primary source ↓e20PDF p. 3, Figure 2 and Sections 2–3.1; p. 9, Section 5 Generative World Models for Robotics
Shared action/future-token DiT and direct action denoising establish the mechanism. The authors distinguish policy/world-model co-training from planning, naming planning as a possible future extension.
Go to primary source ↓8.5 Primary sources
FLARE: Robot Learning with Implicit World Modeling ↗
PDF · 9,649 extracted words
Source fingerprint
37245f9a586229ed477b1c5194b051973ba5b3f050d25cc76e8aefe6a2845aeb