Pith.

Robotics · arXiv:2608.22035 v1 · submitted 22 Aug 2026 · 17 references

Ludi${}_{\scriptscriptstyle 0.1}$: An Agentic System for Socially Intelligent Robots

Wooseong Chung, William Cong, Jakub Dworakowski, Ethan Ewer, Tri Wahyu Guntara, Yeonwoo Jeong, Tianchong Jiang, Chaewon Kim

show 9 more authors

Hyunseo Kim, Jinwoo Kim, Jinyeon Kim, Yea-Seul Kim, Jack Kunde, Kangwook Lee, Sangheon Lee, Robert Nowak, Junha Roh

The paper says

“Robot foundation models have substantially advanced perception and control, but natural human-robot collaboration requires more than executing isolated commands. A robot must recognize ambiguity, maintain context across turns, communicate its intentions, and revise ongoing behavior as the user's intent changes.”

more of the abstract

We present $\scriptstyle\mathsf{Ludi}_{\scriptscriptstyle 0.1}$, an agentic system for socially intelligent robots that integrates interactive speech, multimodal reasoning, memory, navigation, and learned manipulation. Its decision-making core is a fine-tuned vision-language model trained on multi-turn interaction traces spanning ambiguous requests, clarifications, corrections, interruptions, mixed social and task dialogue, and multi-step tasks. A purpose-built harness manages the model-tool interaction loop, while specialized navigation and manipulation policies execute physical skills. Ludi${}_{\scriptscriptstyle 0.1}$ demonstrates a practical path toward fluid human-robot collaboration today while producing the multimodal interaction traces needed to develop a more deeply integrated foundation model for robots and people.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 27 Aug 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A credible system-integration report whose headline numbers rest on a self-referential benchmark; worth engaging for the engineering, but the social-intelligence claim needs external validation. This line is wrong

3 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 17 references say.

The paper says, at §3.2, §3.4, Table 1

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The central claim that fine-tuning roughly doubles success on Interaction Core28 is threatened by a potential circularity between the training-data generator and the benchmark's gold trajectories. Section 3.2 states that 'a state machine and policy table determined every tool call, its arguments, and its result' in the 14,432 training traces, while Section 3.4 defines success via 'a gold tool-call trajectory' and deterministic interaction gates.

more of objection 1

The paper never establishes that the Core28 gold trajectories and gates were authored independently of that same policy table, and no release of scenarios, checkpoints, or trajectories allows the reader to audit the overlap. If the benchmark encodes the same behavioral conventions as the training generator, then the reported gains measure how well the fine-tuned model imitates that generator, not socially intelligent generalization to new interaction patterns. Section 8's honest listing of 'context loss, latency, brittle handoffs' does not address this construct-validity concern. The authors should either (i) provide an external benchmark authored independently of their generator, (ii) show that Core28 scenarios were authored blind to the training policy table and release the full scenario specifications and gates, or (iii) at minimum add a head-to-head comparison on an existing HRI benchmark such as HELPER-X's evaluation suite, and temper the 'practical path' claim to what the evidence supports.

The paper says, at Table 1

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

Table 1 reports a single run per model without error bars, confidence intervals, or information about decoding stochasticity. With only 28 scenarios, the reported differences of 7 to 9 successes are not statistically demonstrated; a single changed scenario outcome could alter the headline conclusion.

more of objection 2

The authors should report results over multiple seeds or rollout repetitions of each scenario, and provide a statistical test (e.g., a paired or bootstrap test) for the fine-tuning gain. Without this, the 'roughly doubles' claim is not quantitatively supported at the reported precision.

The paper says, at §3.4, Table 1

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The LLM judge used to overturn language-gate failures is a potential source of evaluation inflation. Table 1 shows that each fine-tuned model receives exactly one judge-overturned success, which is small, but the judge's reliability is not assessed: there is no inter-annotator agreement, no manual verification of the overturned episodes, and no example transcripts of judge decisions.

more of objection 3

Since the judge is instructed to rule on 'semantic content of a spoken utterance,' it could be lenient toward paraphrases that a human experimenter would reject. The authors should report the judge prompt, show the overturned episodes in an appendix, and provide human agreement on a random sample of gate failures.

6 minor objections

The paper says, at Abstract and §1

Pith says: minor objection 1 of 6

The abstract and introduction present the real-robot Coke/Pepsi episode as a demonstration of the system, but this is a single anecdotal video; the paper would benefit from clearly labeling this as an illustrative demonstration rather than a measured result, and from reporting whether the episode was scripted or fully autonomous.

The paper says, at §3.2

Pith says: minor objection 2 of 6

The fine-tuning details are incomplete: no information is given about LoRA rank, learning rate, number of epochs, sequence length, or the teacher LLM used for language generation. This limits reproducibility.

show 4 more

The paper says, at Figure 3

Pith says: minor objection 3 of 6

Figure 3's caption is confusing: it mentions 'IMG + TURN' and lists 'OBS FRONT/RIGHT/BACK/LEFT' but the schematic is not fully legible, especially the relationship between the four-view observation and the 'middle laptop' grounding. A clearer annotation or a larger figure would help.

The paper says, at §4

Pith says: minor objection 4 of 6

The speech processing numbers (94.7 ms average STT) are presented as a system benchmark, but the text notes they are not end-to-end latencies; the authors should also report perceived response latency or the end-to-end turnaround for a simple query to avoid over-interpretation.

The paper says, at §6

Pith says: minor objection 5 of 6

The VLA section reports simulation calibration metrics but not real-robot manipulation success rates; adding success counts for the physical G1 pick-and-place tasks would strengthen the claim that the specialized policies are functional.

The paper says, at General

Pith says: minor objection 6 of 6

There are minor typographical inconsistencies in notation, e.g., 'Ludi${}_{\scriptscriptstyle 0.1}$' in the abstract versus plain 'Ludi0.1' in the body; these should be unified.

What would settle it

Run an external held-out validation: have researchers who did not see Ludi's state machine, policy table, or training data author 28 new scenarios with gold trajectories and deterministic gates derived from real human-robot interaction transcripts, then evaluate the released vanilla and fine-tuned Qwen3.5-4B/9B checkpoints on this set under the same harness and AI2-THOR environment. If the fine-tuned advantage shrinks materially relative to 17/28 vs 10/28 (4B) or 20/28 vs 11/28 (9B), the Core28 doubling reflects generator circularity rather than generalizable social intelligence.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 27 Aug 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 1 s · 10,818 tokens read, 5,869 written
  3. Editor wrote the verdict above 20 s · 11,630 tokens read, 2,180 written
  4. Referee major revision: 3 major objections, 6 minor 31 s · 11,081 tokens read, 2,945 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Here's the quick read. Ludi0.1 is a genuine system-building effort: fine-tuned Qwen3.5 VLM as a tool-using reasoning core, harness for real-time speech, keypoint navigation, and GR00T-based VLA manipulation, all integrated on a Unitree G1. The new bits are the synthetic multi-turn interaction corpus (14k trajectories covering ambiguity, corrections, interruptions) and the closed-loop Core28 benchmark. The one real-robot demonstration is credible, and Section 8 is refreshingly honest about the operational limits.

The problem is that the central quantitative claim—fine-tuning roughly doubles success on Core28 (10/28 to 17/28 for 4B, 11/28 to 20/28 for 9B)—rests on a benchmark that appears to encode the same behavioral policy used to generate the training data. Section 3.2 says a state machine and policy table determined every tool call and argument in training traces; Section 3.4 defines Core28 success via gold tool-call trajectories and deterministic gates. The paper never shows the gold trajectories were authored independently of that policy table. So the benchmark may be testing how well the model imitates the generator, not transferable social intelligence. That's not a small caveat; it's the load-bearing premise for the 'practical path toward fluid human-robot collaboration today' claim.

Other soft spots are proportionate. Single runs, no error bars (28 scenarios, so noise is nontrivial). No release of scenarios, checkpoints, or trajectories, so no one can audit the overlap. The LLM judge can overturn language-gate failures, but only one judge-overturn per row, so that's minor. The full evaluation is simulation-only; real evidence is one video.

What the paper does well: the engineering is real, the related work is properly cited (Inner Monologue, Hi Robot, Gemini Robotics ER 2), and the limitations discussion is honest. The calibrated Isaac Sim for VLA policy screening is a nice practical touch.

My bottom line: this is a serious system paper with a construct-validity problem in its evaluation. It deserves peer review, but the reviewers should push for either an externally authored held-out benchmark or a release that lets the community check the training/eval overlap. I wouldn't cite it for its numerical claims, but I'd point to it as an example of the current agentic-integration approach.

The strongest objection

Section 3.2 states that the training traces were produced by "a state machine and policy table [that] determined every tool call, its arguments, and its result," with a teacher LLM writing only the language. Section 3.4 defines Core28 success via "a gold tool-call trajectory" plus deterministic interaction gates and a final simulator state. The paper never establishes that Core28's gold trajectories and gates were authored independently of that same policy table, and no release of scenarios, checkpoints, or trajectories lets the reader audit the overlap. If the gold trajectories encode the same conventions as the training generator, then Core28 measures how well the fine-tuned model imitates the generator, not socially intelligent generalization. The reported gains (10/28 to 17/28 for 4B; 11/28 to 20/28 for 9B) could then be a training/evaluation circularity artifact. The LLM-judge appeal is not the main risk: Table 1 shows only one judge-overturned episode per row. The main risk is construct validity of the benchmark itself, which is the exact premise needed for the paper's "practical path toward fluid human-robot collaboration today" claim. Section 8 honestly lists operational limits, but those limits do not address benchmark circularity.

The reader's weakest assumption — that synthetic traces are representative and Core28 measures socially intelligent behavior — is correct, and the most load-bearing version of that concern is sharper: the training-data generator and the evaluation benchmark may share the same state-machine and policy-table conventions. This is not a question of whether the numbers are accurately reported; it is a question of what they measure. If the gold trajectories in Core28 were authored under the same policy table that generated the 14,432 training traces, then fine-tuning's apparent doubling of success is partly or wholly an artifact of the model learning to match its own training generator. The paper provides independent supporting evidence — a real Unitree G1 demonstration, calibrated Isaac Sim screening, and an explicit limitations section — but none of that addresses the benchmark's construct validity. The reader's CONDITIONAL verdict remains appropriate: the architecture is plausible and honestly scoped, but the headline quantitative claim requires external validation before it can support the broad practical-path assertion. I keep the verdict unchanged rather than moving to reject, because the paper's own framing is modest and the identified concern is testable rather than demonstrably fatal.

The referee's report

If the central claims hold, the paper would be a useful systems contribution: it demonstrates an integrated architecture that combines a fine-tuned VLM reasoner, a purpose-built harness, speech, navigation, and VLA policies, and it introduces a closed-loop evaluation methodology for multi-turn human-robot interaction. The acknowledgement of limitations in Section 8 and the decision to deploy the smaller 4B model for latency reasons are honest engineering choices. However, the significance of the quantitative results depends entirely on whether Core28 measures socially intelligent generalization rather than imitation of the synthetic data generator; the current evidence does not establish this, and the paper provides no external validation, no error bars, and no release of scenarios or checkpoints for independent audit. The real-robot evidence is limited to one demonstration. These issues are load-bearing because the abstract and Section 2 claim a 'practical path toward fluid human-robot collaboration today.'

Pith's reading of the paper (a machine's summary, not the authors')

Ludi0.1 is an agentic system for humanoid robots that integrates speech, multimodal perception, memory, navigation, and learned manipulation under a single reasoning core: a vision-language model fine-tuned on synthetic multi-turn interaction traces. The paper's central empirical claim is that this fine-tuning roughly doubles end-to-end task success on a new closed-loop benchmark, Interaction Core28: the 4B model improves from 10/28 to 17/28 and the 9B model from 11/28 to 20/28, approaching the 26/28 of a frontier model with an optimized prompt. The authors argue this is a practical path toward fluid human-robot collaboration today, since the whole system runs on an onsite GPU without cloud inference. The same interaction traces are meant to serve as training data for a more deeply integrated robot foundation model in the future.

What carries the argument. The load-bearing component is the fine-tuned VLM as the agent's decision-maker, which sees the ego view, a 360-degree panoramic context, transcribed speech, interaction history, and tool results, and selects one action from a fixed tool set: say, lookup_keypoints, navigate, reachability, vla, and wait. Surrounding it is a harness that enforces contracts (at most one action-bearing tool per turn, reachability before a pick), executes the chosen tool, records the outcome, and loops back to the model. The training data are synthetic multi-turn traces produced by a state machine and policy table, with a teacher LLM writing the language and rule-based and LLM-judge quality gates filtering each turn; the evaluation runs in AI2-THOR with ProcTHOR-generated homes, scoring both deterministic interaction gates and the final simulator state.

Core claim. The discovery is that fine-tuning a general vision-language model on fully synthetic, multi-turn interaction traces—generated by a state machine and policy table so that every tool call is correct by construction—yields large, measurable gains in socially intelligent robot behavior. On the authors' Interaction Core28 benchmark, which scores the full loop of perception, dialogue, navigation, and manipulation in closed-loop simulation, the fine-tuned Qwen3.5-4B completes 17/28 scenarios versus 10/28 for the vanilla model, and the fine-tuned 9B completes 20/28 versus 11/28, with only one judge-overturned language gate each. The resulting system, Ludi0.1, can handle ambiguous requests (asking a clarification before acting), mid-task corrections (abandoning a reach for Coke and fetching Pepsi instead), interruptions, and mixed social-and-task dialogue on a real Unitree G1 humanoid. The paper frames this as evidence that agentic modular architectures can deliver fluid human-robot collaboration today while producing the interaction traces needed for a more integrated foundation model later.

The premise it rests on. That the synthetic interaction traces—generated from a state machine and policy table—faithfully represent the range and difficulty of real human-robot interactions, and that Core28's deterministic gates and LLM judge measure socially intelligent behavior.

If the paper is right (Pith's inferences, not the authors')

Fine-tuning on synthetic interaction traces can more than double a robot's performance on complex, multi-turn collaborative tasks without changing the base model architecture or the physical skills.

The agentic decomposition—VLM reasoner plus harness plus specialized navigation and manipulation policies—is sufficient to run real-time, interruption-handling interactions on a locally deployed GPU, without cloud inference.

The same synthetic trace pipeline can be scaled combinatorially to cover more tasks, rooms, and interaction events, and the recorded traces can serve as pretraining data for a more deeply integrated robot foundation model (the authors' Ludi 1.0 direction).

Core28-style closed-loop benchmarks, which score the full interaction loop rather than isolated predictions, can reveal gains in social competence that static VQA-style benchmarks would miss.

If the synthetic-to-real gap is small, the same recipe—state-machine-generated traces plus fine-tuning—could transfer to other embodiments and other home layouts, because the tool interface abstracts away embodiment-specific motor control.

The reported remaining failures concentrate in spatial grounding at the delivery end of tasks; a natural extension is to add a learned spatial grounding module or a keypoint-tuned reasoner to close that gap.

The benchmark's LLM judge, however, is itself a language model evaluating language; a stricter test would measure user comprehension or task outcome in human trials, which could change the success counts.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Robot foundation models have substantially advanced perception and control, but natural human-robot collaboration requires more than executing isolated commands. A robot must recognize ambiguity, maintain context across turns, communicate its intentions, and revise ongoing behavior as the user's intent changes. We present $\scriptstyle\mathsf{Ludi}_{\scriptscriptstyle 0.1}$, an agentic system for socially intelligent robots that integrates interactive speech, multimodal reasoning, memory, navigation, and learned manipulation. Its decision-making core is a fine-tuned vision-language model trained on multi-turn interaction traces spanning ambiguous requests, clarifications, corrections, interruptions, mixed social and task dialogue, and multi-step tasks. A purpose-built harness manages the model-tool interaction loop, while specialized navigation and manipulation policies execute physical skills. Ludi${}_{\scriptscriptstyle 0.1}$ demonstrates a practical path toward fluid human-robot collaboration today while producing the multimodal interaction traces needed to develop a more deeply integrated foundation model for robots and people.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-23

The authors' main claim, as the classifier picked it out 1.00

“Ludi${}_{\scriptscriptstyle 0.1}$ demonstrates a practical path toward fluid human-robot collaboration today while producing the multimodal interaction traces needed to develop a more deeply integrated foundation model for robots and people.”

How the main claim is stated: asserted 0.99

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.39 · human participants 0.42 · controlled comparison 0.02 · randomized assignment 0.01 · states quantitative result 0.05 · compares with prior results 0.08 · mentions code or data availability 0.05

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Robot foundation models have substantially advanced perception and control, but natural human-robot collaboration requires more than executing isolated commands. background 0.87 0.05 0.05 asserted 1.00
A robot must recognize ambiguity, maintain context across turns, communicate its intentions, and revise ongoing behavior as the user's intent changes. background 0.55 0.06 0.05 asserted 1.00
We present $\scriptstyle\mathsf{Ludi}_{\scriptscriptstyle 0.1}$, an agentic system for socially intelligent robots that integrates interactive speech, multimodal reasoning, memory, navigation, and learned manipulation. contribution 1.00 0.92 0.14 asserted 1.00
Its decision-making core is a fine-tuned vision-language model trained on multi-turn interaction traces spanning ambiguous requests, clarifications, corrections, interruptions, mixed social and task dialogue, and multi-step tasks. method 0.99 0.89 0.06 asserted 1.00
A purpose-built harness manages the model-tool interaction loop, while specialized navigation and manipulation policies execute physical skills. method 0.99 0.78 0.04 asserted 1.00
Ludi${}_{\scriptscriptstyle 0.1}$ demonstrates a practical path toward fluid human-robot collaboration today while producing the multimodal interaction traces needed to develop a more deeply integrated foundation model for robots and people. interpretation 0.94 0.79 0.83 asserted 0.99

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.85

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 5,038 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

References (17)
  1. [1] L. A. Agrawal, S. Tan, D. Soylu, N. Ziems, R. Khare, K. Opsahl-Ong, A. Singhvi, H. Shandilya, M. J. Ryan, M. Jiang, et al. GEPA: Reflective prompt evolution can outperform reinforcement learning. InInternational Conference on Learning Representations, volume 2026, pages 8479–8565, on Pith
  2. [6] A.-C. Cheng, Y. Ji, Z. Yang, Z. Gongye, X. Zou, J. Kautz, E. Bıyık, H. Yin, S. Liu, and X. Wang. NaVILA: Legged robot vision-language-action model for navigation.arXiv preprint arXiv:2412.04453, on PithPith's reviewarXiv
  3. [8] URL https://openreview.net/forum?id=nZeVKeeFYf9. W. Huang, F. Xia, T. Xiao, et al. Inner monologue: Embodied reasoning through planning with language models.arXiv preprint arXiv:2207.05608, on PithPith's reviewarXiv
  4. [9] C. Y. Kim, C. P. Lee, and B. Mutlu. Understanding large-language model (llm)-powered human-robot interaction. InProceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction, pages 371–380, 2024a. doi: 10.1145/3610977.3634966. M. J. Kim, K. Pertsch, S. K... on PitharXiv
  5. [10] URL https://www.krafton.ai/blog/posts/2026-04-15-pubg_ally_nemotron/ pubg-ally-nemotron-en.html. Z. Lei, C. Liu, Y. Xiong, M. Xiong, Y. Ding, Z. Zhang, W. Li, and S. Chen. Towards long-horizon embodied agents with tool-aligned vision-language-action models.arXiv preprint arXiv... on PithPith's reviewarXiv
  6. [11] J. Liang, W. Huang, F. Xia, P. Xu, K. Hausman, B. Ichter, P. Florence, and A. Zeng. Code as policies: Language model programs for embodied control.arXiv preprint arXiv:2209.07753, on PithPith's reviewarXiv
  7. [12] H. Liu, A. Chen, Y. Zhu, A. Swaminathan, A. Kolobov, and C.-A. Cheng. Interactive robot learning from verbal correction.arXiv preprint arXiv:2310.17555, 2023a. H. Liu, C. Li, Q. Wu, and Y. J. Lee. Visual instruction tuning.Advances in Neural Information Processing Systems, 36:... on PithPith's reviewarXiv
  8. [15] doi: 10.1109/LRA.2023.3236571. Z. Yang, Y. Chen, X. Zhou, J. Yan, D. Song, Y. Liu, Y. Li, Y. Zhang, P. Zhou, H. Chen, et al. Agentic Robot: A brain-inspired framework for vision-language-action models in embodied agents.arXiv preprint arXiv:2505.23450, on PitharXiv
  9. [16] S. Ye, Y. Ge, K. Zheng, et al. World action models are zero-shot policies.arXiv preprint arXiv:2602.15922, on PithPith's reviewarXiv
  10. [17] H. Zhen, X. Qiu, P. Chen, J. Yang, X. Yan, Y. Du, Y. Hong, and C. Gan. 3D-VLA: A 3d vision-language-action generative world model.arXiv preprint arXiv:2403.09631, on PithPith's reviewarXiv
  11. [2019] doi: 10.1145/3292500.3330701. J.-B. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, et al. Flamingo: a visual language model for few-shot learning.Advances in Neural Information Processing Systems (NeurIPS), 35:23716–23736, on PitharXiv
  12. [2021] A. Z. Ren, A. Dixit, A. Bodrova, et al. Robots that ask for help: Uncertainty alignment for large language model planners.arXiv preprint arXiv:2307.01928, on PithPith's reviewarXiv
  13. [2022] D. Driess, F. Xia, M. S. M. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, W. Huang, Y. Chebotar, P. Sermanet, D. Duckworth, S. Levine, V. Vanhoucke, K. Hausman, M. Toussaint, K. Greff, A. Zeng, I. Mordatch, and P. Florence. PaLM-E: An embod... on PithPith's reviewarXiv
  14. [2023] G. Sarch, S. Somani, R. Kapoor, M. J. Tarr, and K. Fragkiadaki. HELPER-X: A unified instructable embodied agent to tackle four interactive vision-language domains with memory-augmented language models.arXiv preprint arXiv:2404.19065, on PithPith's reviewarXiv
  15. [2024] doi: 10.15607/RSS.2024.XX.049. J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y. Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, on PithPith's reviewarXiv
  16. [2025] K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, J. Tanner, Q. Vuong, A. Walling, H. Wang, and U. Zhilinsky.𝜋0: A vision-la... on PithPith's reviewarXiv
  17. [2026] M. Ahn, A. Brohan, N. Brown, et al. Do as I can, not as I say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, on PithPith's reviewarXiv
Cited by 1 Pith paper

Reviewed papers on Pith that cite this one. The full citation record

  1. PUBG Ally: A Conversational Embodied Agent as an AI Teammate cs.AI · Sep 2026
How the literature uses it

What this paper does with the arXiv papers it cites

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

What it cites · 2 sentences without a confident label

Uses the method or tool of 2 papers
Cites as background 11 papers
  • Models built on pretrained large language and vision–language models exhibit increasingly general perception, reasoning, and language-conditioned control across navigation and manipulation [ref] , while more recent approaches have begun to incorporate generative world models to support prediction and action generation [ref] .

    NaVILA: Legged Robot Vision-Language-Action Model for Navigation · Introduction · 0.93

  • Models built on pretrained large language and vision–language models exhibit increasingly general perception, reasoning, and language-conditioned control across navigation and manipulation [ref] , while more recent approaches have begun to incorporate generative world models to support prediction and action generation [ref] .

    World Action Models are Zero-shot Policies · Introduction · 0.93

  • Models built on pretrained large language and vision–language models exhibit increasingly general perception, reasoning, and language-conditioned control across navigation and manipulation [ref] , while more recent approaches have begun to incorporate generative world models to support prediction and action generation [ref] .

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots · Introduction · 0.93

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment