Embodied-R1: Points as a Cross-Embodiment Interface—Reinforced Reasoning, Evidence, and Reproducibility

0. Why it matters, and how not to interpret it

The appeal is not a generic 87.5% success rate, but four programmatically scorable, cross-arm pointing capabilities: REG (referred object), RRG (placement region), OFG (functional part), VTG (visual trajectory). This recasts the difficult interface between language and control as 2D points/short paths, trained with staged GRPO and task-composed rule rewards. Versus same-data SFT, Table 2 gains on RoboRefIt / Where2Place / VABench-P / Part-Afford are 1.73 / 28.25 / 15.54 / 16.43 pp(Table 2, p.8; recomputed). RL helps most on open-space, part, and trajectory tasks with multiple valid answers and unnatural string supervision, rather than simple object localization.

External action evidence is less clear-cut than the abstract suggests. SIMPLEREnv average 56.2% exceeds strongest baseline SoFar's 53.8% by only 2.4 pp / 4.46% relative, and is not best on two of four tasks (Table 4, p.10). Real-robot 87.5% comes from 46 trials across eight tasks, averaged by task (Table 5, p.11), without CIs. With five or six trials each, one outcome changes scores by 16.7–20 pp. More seriously, the main-text RRG reward is 0.1 format + 0.2 distance + 0.7 mask(p.7), while Appendix Table 9 gives 0.1 + 0.3 + 0.6(p.24). The defensible conclusion is:point-based intermediate interfaces plus verifiable RL help connect perception to action, but evidence does not establish universal VLA superiority or open-world success at 87.5%.

1. Link verification table

Search boundaries. This round checked arXiv, ICLR/OpenReview, RSS/CoRL, Hugging Face Papers, Semantic Scholar, Papers with Code, and ACL Anthology. ACL yielded no more direct candidates than robotics/ICLR sources. Some Papers with Code entries were unstable, so they served only identity checks. All numbers below come from six PDFs, not search snippets, leaderboards, or project pages.

2. Candidate pool, deduplication, and random selection

First read ml-report-index.md. The past 14 days covered DreamGen (world-model robot trajectories, 09-03) and VIKI-R (multi-robot rewards, 09-10), excluding those neighboring themes. Today's scope retains single-robot zero-shot manipulation, embodied reasoning, and agent RL. Proceedings/OpenReview, arXiv, and Hugging Face Daily Papers supplied 24 candidates, checked through Semantic Scholar/Papers with Code; all core candidates are within 24 months.

Fixed N=24 then executed shuf -i 1-24 -n 1returned 20. Embodied-R1 has formal ICLR 2026 status, a 31-page PDF/appendix, open models/data/code, and real-robot results, providing sufficient material, sono redraw was needed. Selection preceded deep reading, without ranking by popularity or subjective interest.

3. Motivation: what is the perception–action gap? {#motivation-gap}

End-to-end VLAs map images/instructions directly to robot-specific action tokens. Large trajectory corpora enable closed-loop control, but outputs depend on arm, camera, control rate, and joints; new embodiments often require data collection/fine-tuning. OpenVLA-OFT shows strong specialized adaptation: average LIBERO success rises from 76.5% to 97.1%and action-generation throughput increases 26×(Table I, p.6). This is target-robot/task fine-tuning, not action-data-free cross-embodiment zero-shot transfer.

Modular VLM-plus-planner systems separate semantics/control. RoboPoint uses synthetic data to predict feasible image points conditioned on actions, improving spatial-affordance accuracy over GPT-4o/PIVOT by 21.8%and downstream success by 30.5%(abstract/experimental summary, pp.1,10). Its points mainly address REG/RRG, not systematic functional parts or continuous paths. FSD adds Spatial-relationship CoT and visual aids, reaching SIMPLEREnv 40.6%and eight-real-task success 72%(abstract, Table 4/real experiments, pp.1,11), but remains primarily SFT-based, where one labeled point can penalize other valid answers.

Embodied-R1 identifies four specific issues: scarce heterogeneous action data; language too coarse and joint actions too specific; multiple valid points/paths penalized by cross-entropy when unlabeled; and spatial reasoning chains needing consistency with final points. It chooses pointing as an intermediate bandwidth: more geometric than language, more general than actions, and automatically rewardable through masks, distances, and trajectory RMSE.

4. Narrative, assumptions, and tradeoffs

Four steps define the system. REG predicts a point inside the manipulated object; RRG inside a relationally valid placement region; OFG on a graspable/pressable/pullable part; VTG predicts eight object-center path points. Existing VQA, spatial, and robotics data form a two-stage curriculum: general spatial reasoning, then point-centric specialization. GRPO updates grouped image–instruction responses with relative rewards, avoiding a critic. At inference, points/paths are back-projected into 3D and executed through grasping, IK, path planning, or Diffusion Policy.

Four assumptions underlie this:Geometric sufficiency—2D points capture most pre-action decisions;Scoring sufficiency—masks, distances, and trajectory RMSE approximate real action value;Transferable decomposition—after perception/reasoning transfers, low-level planners are not the main bottleneck;Acceptable open-loop planning—plans from one image/short sequence do not rapidly fail under occlusion, contact, or moving objects. Benefits are interpretability, cross-embodiment use, and no task-action fine-tuning; points cannot fully express contact forces, twisting, wiping, deformables, or long-horizon correction.

5. Equation-level formalization and algorithm

Input: Qwen2.5-VL-3B-Instruct, Stage-1 spatial/general data,
       Stage-2 Embodied-Points tasks, task reward library

Stage 1: 2 epochs RFT
  for image, instruction in Embodied-Spatial-84K + ViRL subset(18K):
    sample 8 responses; score format + exact/spatial correctness
    normalize group rewards; GRPO update with KL to reference

Stage 2: 1 epoch RFT
  for task in {REG, RRG, 3D-RRG, OFG, VTG, QA}:
    sample 8 <think> + coordinate responses
    choose task-specific rewards {format, acc, mask, distance, trace, env}
    GRPO update

Deployment:
  RGB(/D) + instruction -> reasoning + point(s)/8-point trace
  -> pixel-to-world / detector / IK / motion planner (or DP)
  -> robot executes; no task-specific fine-tuning in zero-shot setting

6. Data construction, engineering, and reproducibility

Qwen2.5-VL already uses native dynamic-resolution ViT, Window Attention, and variable visual tokens (technical report §2, Table 1, pp.3–4). Point grounding is inherited, not created from nothing. Embodied-R1 reorganizes it into embodied tasks/data/rewards. Without a matched-size backbone lacking native grounding, one cannot claim any 3B VLM would gain equally from GRPO.

7. Benchmarks and metrics: what numbers measure

8. Main results: perception-to-execution matrices

9. Ablations, sensitivity, and transitions

10. Statistical significance and cost

Apart from fixed episode counts in some experiments, main tables lack CIs, standard deviations, multiple seeds, or significance tests. At 24 SIMPLEREnv episodes/task, even an approximate independent-binomial view requires raw episodes to judge 56.2% versus 53.8%. Real tasks with five/six trials are weaker still: one trial changes 16.7–20 pp. Larger pointing benchmarks still omit training randomness.

Training is approximately 8×A100×(48+48)h = 768 A100 GPU-hour(recomputed from Appendix p.23), excluding GPT-4o generation, Grounded-SAM, CoTracker3, baselines, and repeated trials. Dollar costs, energy, generation throughput, and total wall-clock are absent. Execution time is given for only two tasks: R1 10s/8s, OpenVLA-FT 23s/12s, MOKA 28s/18s, FSD 14s/10s (Table 10, p.25), without hardware, variance, or stage latency. This supports faster demonstrations, not a universal serving curve.

11. Genuine contributions, reusable design, and transfer potential

The first contribution isinterface design: points, regions, functional parts, and short trajectories unify grounding, affordances, placement, and traces. Second,a composable reward librarycombines format, exact, mask, distance, trace, and environment rewards by task, scaling better than bespoke losses. Third,a spatial-first, action-second RFT curriculumbuilds semantics through general spatial questions before aligning embodied point outputs. Fourth,a bridge from modular planning to learned policies: Table 12 shows point priors can condition or prompt Diffusion Policy rather than remain tied to conventional planners.

For GUI agents, REG maps to controls, RRG to drop zones, and VTG to cursor paths. Surgery/industry can map OFG to safe contact regions but need uncertainty/force constraints. Navigation can use points/curves as waypoints, extending single images to spatiotemporal maps. Data selection can prioritize group-reward variance as remaining learning signal. Direct transfer is unsuitable for sustained contact forces, deformation, or hidden-state estimation because 2D rewards lack sufficient dynamics information.

12. Technical lineage: problems solved, gaps, responses, and transferred designs

General VLM backbone (Qwen2.5-VL: dynamic resolution, native points/boxes)
          │
          ├──────────────► End-to-end actions
          │                 OpenVLA ──► OFT: chunking + continuous actions + L1
          │                                      │
          │                                      └── high in-task success, adaptation data required
          │
          └──► Embodiment-independent representations
                RoboPoint: REG / RRG points
                    │
                    ├──► FSD: Sr-CoT + visual aids (SFT)
                    │        │
DeepSeekMath: GRPO ─┴────────┴──► Embodied-R1
                                     ├─ REG / RRG / OFG / VTG
                                     ├─ mask / distance / trace rewards
                                     ├─ zero-shot planner
                                     └─► DP + point prior (back into closed-loop policies)

13. Author-stated limitations

1. Limited policy integration. The authors primarily connect to traditional motion planners, with only preliminary DP validation; deeper learned-policy integration is future work (Limitations p.29). Main results are not unified end-to-end system comparisons.
2. Limited long-horizon tasks and expressiveness. Instructions are mostly single-step without automatic decomposition. Points poorly express force control, twisting, wiping, and deformable manipulation (p.29). General manipulation means generalization within current task families, not open-world sufficiency.
3. Fragile 3D and physical common sense. RGB-D experiments exhibit depth hallucination and slightly worse complex relations; intermediate reasoning can contradict physics (p.29).

14. Independent critique: evidence-to-inference chains

1. [Attribution] Because R1 improves same-data SFT by only 1.73 ppon RoboRefIt but by 28.25 ppon Where2Place (Table 2, p.8),therefore evidence favors RL for multivalued geometry, not uniform gains across grounding. Report results by task reward entropy.
2. [Statistics] Because real tasks have five/six trials each, 46 total, without CIs (Table 5, p.11),therefore 87.5% is not a stable population success estimate; one outcome changes a task by 16.7–20 pp.
3. [External validity] Because SIMPLEREnv Stack trails SoFar by 26.4 ppand Eggplant trails OpenVLA-OFT by 14.2 pp(Table 4, p.10),therefore the 56.2% mean hides task heterogeneity and cannot mean universal SOTA.
4. [Internal consistency] Because main text and Table 9 give RRG mask/distance weights 0.7/0.2 versus 0.6/0.3 (pp.7,24),therefore the key training configuration is ambiguous; reproduction without code checks may differ.
5. [Recomputability] Because Table 4's strongest mean baseline is 53.8% versus R1 56.2%, only 4.46% relative improvement, while real-table baselines move 25.0%→87.5%,therefore the abstract's 62% improvement lacks a clear denominator in main tables and remains an underspecified aggregate claim.
6. [Reward proxies] Because removing eight-point constraints yields only endpoints and raises RMSE 77.83→105.2 (Table 11, p.27),therefore models exploit scoring gaps. A point inside a mask may still be ungraspable; future rewards need collision, reachability, and execution value.
7. [Data definitions] Because the paper says Embodied-Points-200K/3B while HF shows about 321k rows/about 4B parameters,therefore deduplication, stage mixing, and parameter accounting remain unresolved, preventing exact compute/sample-efficiency recomputation.
8. [Comparison fairness] Because modular R1 uses external perception/planning/IK while VLA baselines output actions end to end, without equal hardware, control frequency, or wall-clock (Tables 4,10),therefore success cannot be attributed solely to representations. External-module failure rates and latency belong in system comparisons.

15. Open challenges: executable head-to-head gaps

16. Recent-work radar (not sources for core numbers)

1. SimpleVLA‑RL: applies RL directly to VLA policies, suitable for compute-matched closed-loop comparison with pointing-then-control.
2. ConRFT: emphasizes embodied reasoning/output consistency, matching the finding that thinking needs reward calibration.
3. π0.5: open-world VLA generalization provides stronger end-to-end priors, but evaluation cost/adaptation accounting must align.
4. NVIDIA GR00T N1: humanoid foundation models expand action complexity, testing whether point interfaces remain sufficient.
5. SmolVLA: small open VLA, a closer parameter/deployment-cost end-to-end comparator.
6. SpatialVLA: explicit spatial representations can separate reward gains from backbone spatial inductive biases.
7. OpenVLA‑OFT: strong RSS 2025 adaptation baseline; 97.1% LIBERO and 26× throughput emphasize different goals of zero-shot modularity and in-task performance.
8. FSD: direct ICLR 2026 precursor; SIMPLER 40.6%→R1 56.2% illustrates SFT→RFT evolution, not a strict equal-data/compute causal experiment.

17. Final assessment and reproduction blueprint

Embodied-R1 abstracts pre-action decisions into shared, automatically scorable geometry. RL can align reasoning/points on large image–instruction datasets without expensive real episodes. Tables 6/11 show thinking calibration and structural constraints suppressing reward hacking—more reusable than the slogan of a 3B model at 87.5%.

A credible reproduction should lock arXiv v2/repository commit, record HF snapshots and deduplicated row counts, reproduce both stages, verify RRG weights from code, and run minimal three-point sweeps of eight-point constraints, group=8, and KL=1e-2 . Reproduce Tables 2,6,11 before SIMPLEREnv, then use at least 30–50 real trials/task. System superiority requires R1+planner to reliably beat FSD/RoboPoint at equal hardware/latency and approach OpenVLA-OFT on new embodiments with substantially less adaptation data.

Footnote: artifacts, index, and audit status

• ml-report-index.md This issue's entry has been appended.
• The Warm Editorial single-file HTML uses inline CSS/JS, system-ui plus PingFang SC/Microsoft YaHei, top-right theme switching, result tables, and a single-axis chart with hover values.
• Playwright loads, but /usr/bin/chromium、/usr/bin/chromium-browser、/opt/pw-browsers and its default cache contain no Chromium executable. As instructed, playwright installwas not run, sobrowser screenshot checks were not completed. Warm Editorial validation (0 warnings), HTML structure checks, and responsive mobile table scrolling were completed, but do not replace actual rendering; this remains a delivery limitation.
• This report neither invents a denominator for the unverified 62% nor estimates unreported standard deviations, sensitivity curves, or failure rates from plots.