This assessment is based on version 1 of this paper. Version 2 is now available on arXiv — the authors may have revised their methods, results, or conclusions.
MolmoB0T: Large-Scale Simulation Enables Zero-Shot Manipulation
Abhay Deshpande, Maya Guru, Rose Hendrix, Snehal Jauhri, Ainaz Eftekhar, Rohun Tripathi, Max Argus, Jordi Salvador
Abstract
A prevailing view in robot learning is that simulation alone is not enough; effective sim-to-real transfer is widely believed to require at least some real-world data collection or task-specific fine-tuning to bridge the gap between simulated and physical environments. We challenge that assumption. With sufficiently large-scale and diverse simulated synthetic training data, we show that zero-shot transfer to the real world is not only possible, but effective for both static and mobile manipulation. We introduce MolmoBot-Engine, a fully open-source pipeline for procedural data generation across robots, tasks, and diverse simulated environments in MolmoSpaces. With it, we release MolmoBot-Data, a dataset of 1.8 million expert trajectories for articulated object manipulation and pick-and-place tasks. We train three policy classes: MolmoBot, a Molmo2-based multi-frame vision-language model with a flow-matching action head; MolmoBot-Pi0, which replicates the architecture to enable direct comparison; and MolmoBot-SPOC, a lightweight policy suitable for edge deployment and amenable to RL fine-tuning. We evaluate on two robotic platforms: the Franka FR3 for tabletop manipulation tasks and the Rainbow Robotics RB-Y1 mobile manipulator for door opening, drawer manipulation, cabinet interaction, and mobile pick-and-place. Without any real-world fine-tuning, our policies achieve zero-shot transfer to unseen objects and environments. On tabletop pick-and-place, MolmoBot achieves a success rate of 79.2% in real world evaluations across 4 settings, outperforming at 39.2%. Our results demonstrate that procedural environment generation combined with diverse articulated assets can produce robust manipulation policies that generalize broadly to the real world. Technical Blog: https://allenai.org/blog/molmobot-robot-manipulation
AI Impact Assessments
(3 models)Scientific Impact Assessment: MolmoBot – Large-Scale Simulation Enables Zero-Shot Manipulation
1. Core Contribution
MolmoBot's central claim is that with sufficiently large-scale and diverse simulation data, zero-shot sim-to-real transfer for robot manipulation is achievable without any real-world fine-tuning, photorealistic rendering, or explicit domain adaptation. The paper delivers this through three interconnected contributions:
The headline result is compelling: MolmoBot achieves 79.2% success on real-world tabletop pick-and-place across 4 environments versus 39.2% for π₀.5, a state-of-the-art model trained on >10k hours of real-world demonstrations. The controlled comparison via MolmoBot-Pi0 (same architecture as π₀, trained only on sim data) achieving 46.7% versus π₀.5's 39.2% provides strong evidence that the data itself—not just the architecture—drives performance gains.
2. Methodological Rigor
The experimental design has notable strengths:
However, there are methodological concerns:
3. Potential Impact
This work could significantly shift the field's approach to robotics foundation model training:
The practical limitation is that the approach currently only covers rigid-body and simple articulated manipulation. Many real-world applications require dexterous, contact-rich manipulation that current simulators cannot model faithfully.
4. Timeliness & Relevance
This paper arrives at a critical inflection point. The robotics community is grappling with the tension between (a) the demonstrated power of large-scale real data (π₀, Gemini Robotics, GR00T) and (b) the prohibitive cost and limited accessibility of such data. MolmoBot directly addresses this bottleneck. The simultaneous release of SPOC for navigation and MolmoBot for manipulation from the same group establishes a coherent thesis: scaled simulation can replace real data for foundational robot skills.
The timing also aligns with rapid advances in VLA models, where the data question is increasingly the bottleneck rather than architecture design. The MolmoBot-Pi0 comparison elegantly demonstrates this.
5. Strengths & Limitations
Key Strengths:
Notable Limitations:
Overall Assessment
MolmoBot represents a significant empirical contribution demonstrating that scaled simulation data can produce manipulation policies competitive with or superior to those trained on expensive real-world demonstrations. The open-source commitment amplifies its impact substantially. The architectural innovation (layerwise coupling) and controlled baselines strengthen the scientific claims. However, the scope of manipulation tasks is limited, and the mobile manipulation results are preliminary. The paper's greatest impact will likely be as infrastructure (the engine and data) rather than as a definitive answer to the sim-to-real question.
Generated Mar 18, 2026
Comparison History (234)
Paper 2 likely has higher scientific impact: it challenges a core assumption in robot learning (that sim-only training cannot yield strong zero-shot real-world manipulation) and supports it with large-scale data, open-source generation pipeline, and a sizable released dataset (1.8M trajectories). The approach is broadly applicable across robots and tasks, ties into timely VLM-based robotics, and enables practical deployment without costly real-world data collection. Paper 1 is strong and novel for agile locomotion, but its impact is narrower (humanoid parkour) and less likely to generalize across robotics subfields.
MolmoBot demonstrates stronger scientific impact through its comprehensive challenge to a fundamental assumption in robot learning (that real-world data is necessary for sim-to-real transfer). It provides a fully open-source pipeline, a massive 1.8M trajectory dataset, evaluations across multiple robot platforms and policy architectures, and strong quantitative results (79.2% vs 39.2% over π0.5). The breadth of contribution—data generation engine, dataset, multiple policy classes, and extensive real-world evaluation—combined with open-source availability maximizes community impact. Paper 2's digital teleoperation concept is novel but narrower in scope and less validated comparatively.
Paper 1 has higher potential impact because it directly challenges a widely held sim-to-real assumption and demonstrates strong zero-shot real-world transfer without any real data or fine-tuning, enabled by a large, diverse, fully open-source simulation/data-generation pipeline and a 1.8M-trajectory dataset. This combination is likely to be broadly enabling across robotics labs (lowering data-collection barriers) and applicable to multiple platforms and task families (static + mobile manipulation). Paper 2 is methodologically insightful for VLA training and CoT usage, but its advances are more incremental and may depend more on access to large robot datasets.
Paper 1 has higher likely impact: it challenges a core sim-to-real assumption with large-scale, open-source procedural simulation and a sizable trajectory dataset, demonstrating strong zero-shot real-world performance on multiple robot platforms and tasks. The release of an end-to-end engine + 1.8M-trajectory dataset enables broad follow-on research and real-world deployment, increasing cross-lab reproducibility and adoption. Paper 2 is novel and timely (reward-free test-time RL via confidence), but its primary impact is methodological within VLA adaptation and may be more benchmark-dependent, with less immediate systems-level leverage than Paper 1’s data+infrastructure+real-robot results.
Paper 1 more strongly advances core robot learning by challenging a central assumption (simulation alone cannot yield strong sim-to-real) and substantiating it with large-scale open-source procedural generation plus a 1.8M-trajectory dataset and rigorous zero-shot real-robot evaluations on two platforms with large gains over a named baseline. Its contributions (data engine, dataset, policies) are broadly reusable and likely to catalyze follow-on research across manipulation, sim-to-real, and foundation-policy training. Paper 2 is timely and useful systems engineering for long-horizon autonomy, but appears more integrative and benchmark/task-specific, with less fundamental algorithmic novelty.
Paper 2 has higher estimated impact due to a more novel and broadly applicable methodological contribution: “asymmetric physics” combining high-fidelity non-differentiable simulation with differentiable surrogates to enable scalable end-to-end learning in dense multi-robot, contact-rich settings. It targets swarm robotics—a harder, less-solved regime than single-robot manipulation—and demonstrates scaling to 512 agents plus zero-shot transfer to multiple real quadrupeds, suggesting strong generality. The approach could influence RL, differentiable simulation, multi-agent systems, and locomotion, making its cross-field breadth and timeliness particularly high.
Qwen-RobotManip presents a more comprehensive foundation model approach with a unified alignment framework across multiple dimensions, demonstrated generalization across 15 platforms and multiple benchmarks, and a massive 38,100-hour training corpus. It achieves state-of-the-art across all OOD settings and ranks 1st in RoboChallenge. While MolmoB0T makes a strong contribution showing sim-only zero-shot transfer is viable, Qwen-RobotManip's broader scope—covering cross-embodiment transfer, error recovery, and instruction following—alongside its systematic alignment methodology, positions it for greater impact on the field's trajectory toward general-purpose robotic manipulation foundation models.
MolmoBot challenges the fundamental assumption that real-world data is necessary for sim-to-real transfer, demonstrating effective zero-shot manipulation from purely simulated data at scale. Its fully open-source pipeline, 1.8M trajectory dataset, multiple policy architectures, and strong real-world results (79.2% vs π₀.₅'s 39.2%) make it highly impactful. While EgoInfinity is innovative in converting internet videos to robot actions, MolmoBot's paradigm shift—proving simulation alone suffices—has broader implications for democratizing robot learning and could reshape the field's approach to data collection.
Paper 2 has higher potential impact because it challenges a fundamental assumption in robot learning: that real-world data is required for sim-to-real transfer. By achieving robust zero-shot physical manipulation solely through large-scale synthetic data, it provides a highly scalable paradigm for embodied AI. Furthermore, releasing an open-source procedural generation engine and a 1.8-million trajectory dataset will democratize robotics research and catalyze immense follow-up work. While Paper 1 offers an elegant unified algorithmic approach, Paper 2's empirical breakthrough, superior real-world performance over strong baselines, and massive open-source contributions will likely drive a larger field-wide paradigm shift.
MolmoBot demonstrates a paradigm-shifting result: zero-shot sim-to-real transfer without any real-world data, challenging a fundamental assumption in robot learning. It provides a fully open-source pipeline, a massive 1.8M trajectory dataset, multiple policy architectures, and strong real-world results across two platforms, outperforming π0.5 significantly. Its breadth of impact (open-source tools, large-scale data, multiple robot platforms) and the boldness of its claim—that simulation alone suffices—make it highly impactful. Paper 2 (PRISM) makes solid contributions to memory in visuomotor policies but addresses a narrower problem with more incremental improvements.