When Absolute State… | University of Tokyo, NABLAS… Summary

The One-Line Summary: A deceptively simple fix — redefining a robot's starting position as "zero" at the beginning of each episode — delivers a 15x improvement in task success rates over standard absolute state encoding, and maintains meaningful performance when deployed in conditions the robot has never seen before.

1. Key Themes

Absolute Proprioceptive State Is a Silent Killer for Deployed Robots

The default approach in most imitation learning pipelines — feeding raw joint positions as absolute numbers — catastrophically fails when a robot's starting position varies even slightly. This isn't a niche edge case: it's the standard condition for any robot on a mobile base, linear rail, or adjustable mount. The paper demonstrates this empirically: the absolute encoding (Abs/Abs) achieved only a 5% task success rate in-distribution and 0% out-of-distribution, with the authors noting the OOD evaluation had to be halted entirely due to dangerous robot behavior. As stated in Section V: "During the out-of-distribution evaluation, the policy displayed dangerous movements, prompting us to halt the evaluation and give a 0 score for out-of-distribution." This is not a benchmark failure — this is a safety incident.

Episode-Wise Relative Encoding Is a Practical, Low-Cost Fix With Outsized Returns

The proposed solution requires no new hardware, no external pose estimators, and no architectural changes. It's a coordinate normalization applied at episode start: subtract the robot's initial joint positions from all subsequent states and actions. This single change boosted task success from 5% to 50% in-distribution and 25% out-of-distribution (Table II). The average task score (0–4 scale) jumped from 0.75 to 3.30 in-distribution (Table I, Section V). The authors frame this as: "the episode-wise relative representation effectively collapses the variance in starting conditions, allowing the model to learn more efficiently across all training episodes" (Section VI-B Appendix).

The Train-Test Gap Is a Proprioception Problem, Not Just a Vision Problem

The field has poured resources into visual robustness — invariant features, foundation model encoders, domain randomization. This paper argues the proprioceptive channel has been neglected as a source of distribution shift. As noted in the introduction: "the overwhelming majority of research aimed at closing this train-test distribution gap has focused on the visual domain... proprioceptive inputs, such as joint positions and velocities, are still frequently fed into neural policies as raw, absolute numeric values" (Section I). The implication is that teams spending heavily on visual generalization may be leaving significant performance on the table by ignoring how they encode the robot's own body state.

There Is a Real Performance-Robustness Tradeoff Between Encoding Strategies

The "no state, chunk-wise action" baseline (0/Chunk, analogous to UMI's approach) is more robust to OOD conditions — it actually improved average score by 13% OOD — but it pays a meaningful cost in peak performance (20% success rate vs. 50% for Eps/Eps in-distribution). More critically, it exhibits high variance: "The average score standard deviation is higher than all other policies, showing unreliable performance" (Section V). For operators, this tradeoff is real: stateless chunk-wise policies are safer to deploy in novel conditions but less reliable in known conditions. The episode-wise relative approach threads the needle, though it still degrades 11% OOD.

The Fix Generalizes Beyond This Specific Robot

The authors explicitly flag broader applicability: "While this paper studies a specific task and a specific robot, the results are expected to hold with other robots in other settings that also have linear joints, such as the Agitbot G1, where the torso is set on a vertical rail" (Section VI). Any robot with translational degrees of freedom — humanoids with adjustable torso height, arms on gantry systems, mobile manipulators — faces this exact problem.

2. Contrarian Perspectives

More Data Does Not Solve Coordinate Frame Problems

The conventional scaling playbook says: if your robot fails to generalize, collect more diverse data. This paper challenges that directly. The training dataset is already substantial — 2,500 episodes, ~70 hours at 50Hz, collected across two distinct environments (Section IV-C). Yet the absolute encoding still fails catastrophically. The failure isn't a data volume problem; it's a representational problem. As the introduction states: "Scaling the amount and diversity of the training data has shown some success in improving zero-shot generalization, yet robots still fail when faced with new, unseen test conditions" (Abstract). The implication for teams spending heavily on data collection infrastructure: if your state representation is wrong, no amount of data fixes it.

Removing State Information Entirely Is a Defensible Engineering Choice

Most robotics engineers would resist discarding proprioceptive state — it seems like throwing away useful signal. But the 0/Chunk baseline (no state, chunk-relative actions) actually outperforms absolute encoding by a wide margin (20% vs. 2.5% total success rate, Table II) and shows stronger OOD stability. The paper's framing is direct: "By extending the visual observations available to the policy and using delta end-effector actions, State-Free Policies have demonstrated stronger generalization" (Section II). For teams without the engineering bandwidth to implement episode-wise normalization, simply dropping the state and using chunk-relative actions is a meaningful improvement over the status quo — though with the caveat of higher behavioral variance.

Object-Relative and External Pose Approaches Are Solving the Wrong Problem

Several prominent methods (TAX-Pose, VIOLA) anchor robot actions to external object poses, adding perception pipelines for 6DoF object estimation. This paper argues the same robustness gains can be achieved with purely intrinsic robot state, no external sensors required: "Some works use poses relative to an object in the world, which requires information outside of the robot... Here, we use only the robot's intrinsic state" (Section II). For deployment engineers, this matters because external pose estimators add failure modes, latency, and calibration overhead. A purely intrinsic fix that achieves comparable robustness is operationally superior.

3. Companies Identified

Telexistence Inc.

Description: Japanese robotics company building autonomous robots for retail environments, with a Foundation Model Division.
Why Relevant: The paper originates from Telexistence's Foundation Model Division. The robot used (Ghost) is their commercial platform, and the task (bottle recovery in a convenience store shelf) is their core deployment scenario. This is applied research on a production system, not a lab demo.
Quote: "We reuse the bottle recovery task and the Ghost robot presented in FTACT, but extend the task from a single starting position in front of a table to a variable starting position at the back of a convenience store shelf" (Section IV-A).

Agitbot (referenced, not primary)

Description: Robotics company producing humanoid/mobile manipulator platforms including the G1.
Why Relevant: Explicitly called out as a platform where these findings directly apply: "the results are expected to hold with other robots in other settings that also have linear joints, such as the Agitbot G1, where the torso is set on a vertical rail" (Section VI). The G1's vertical torso rail creates the exact proprioceptive drift problem studied here.

Hello Robot (Stretch)

Description: Maker of the Stretch mobile manipulator, a compact robot designed for indoor/domestic environments.
Why Relevant: Cited as an example of robots with linear rail systems that extend operational workspace — the exact deployment class this paper addresses: "robots are often equipped with mobile bases or linear rail systems to extend their operational workspace" (Section I), with Stretch cited as reference [4].

4. People Identified

Maxime Alvarez

Lab/Institution: Telexistence Inc. Foundation Model Division / University of Tokyo
Why Notable: Lead author. Working at the intersection of academic research and commercial deployment in retail robotics. Dual affiliation suggests research that is designed to ship, not just publish.
Quote: First author on both this paper and the prior FTACT work (cited as [15]), indicating sustained focus on making imitation learning work reliably on Telexistence's platform.

Genki Sano

Lab/Institution: Telexistence Inc.
Why Notable: Senior author and appears on both this paper and FTACT. Likely leads the Foundation Model Division's applied research agenda.
Quote: Listed as corresponding institutional contact for Telexistence's Foundation Model Division.

Ryo Watanabe

Lab/Institution: Telexistence Inc.
Why Notable: Co-author and first author on the prior FTACT paper that established the bottle recovery task baseline this work builds on. Key figure in the continuity of this research program.
Quote: "We reuse the bottle recovery task and the Ghost robot presented in FTACT" — Watanabe et al. 2025 is FTACT [15].

Tony Z. Zhao (cited, not an author)

Lab/Institution: Stanford / Berkeley (ACT paper)
Why Notable: The policy architecture used in this paper (ACT — Action Chunking Transformer) is Zhao et al.'s work [18]. This paper's findings about proprioceptive encoding are directly applicable to the large ecosystem of teams building on ACT.
Quote: "We evaluate our representations using an Action Chunking Transformer (ACT) architecture" (Section I).

5. Operating Insights

Never Deploy a Robot With Linear Joints Using Absolute Proprioceptive State

If your robot has any translational degree of freedom — a torso rail, a gantry, a mobile base contributing to arm workspace — and your policy is trained on absolute joint positions, you are running a system that will fail unpredictably and potentially unsafely when starting conditions vary. The fix costs almost nothing to implement: normalize states and actions relative to episode-start position. The return is a 15–20x improvement in success rate. Any CTO deploying mobile manipulation should audit their state encoding immediately. As the paper demonstrates: "If an end-to-end policy is trained on absolute joint positions, an identical manipulation task executed from a starting position offset by just a few centimeters can result in catastrophic failure or erratic, unsafe movements" (Section I).

When Building Multi-Environment Datasets, State Representation Determines Whether Your Data Combines Cleanly

This paper implicitly surfaces a dataset architecture insight: when data is collected across environments with varying robot starting positions, absolute state encoding means each environment's data looks fundamentally different to the model. Episode-wise relative encoding makes all episodes look the same at time zero, regardless of where the robot started. This is directly relevant for anyone building large-scale robot datasets across multiple sites or operators: "The results suggest a practical path to leveraging data collected by robots with varying frames of reference and deployment to unseen test configurations" (Abstract). Teams planning cross-facility or cross-robot data pooling should encode data in relative frames from the start — retrofitting this later is painful.

6. Overlooked Insights

The Force-Torque Sensor Is Present But Its Contribution Is Uninvestigated

The paper mentions that a 6D force-torque sensor is included in the state, calibrated to zero at episode start: "To that state, we concatenate a 6D force-torque sensor calibrated to 0 at the beginning of the episode" (Section IV-C). This sensor is added to all three conditions being compared, so its contribution is never isolated. For a precision task involving pressing a bottle against a shelf edge (the "Press" stage in Figure 3), force-torque feedback could be load-bearing for performance. The fact that it's always-on means the paper's results may understate how much of the performance comes from the encoding strategy versus the force-torque signal — and teams deploying without F/T sensors can't directly use these numbers to predict their own outcomes.

The OOD Improvement Shown by the Stateless Policy May Be a Measurement Artifact

The 0/Chunk (no state) policy showed a counterintuitive 13% improvement in average score when moved from in-distribution to out-of-distribution conditions. The paper notes this but doesn't fully explain it: "this policy does not suffer any performance loss from the out-of-distribution cases, even improving the average score (+13%)" (Section V). One likely explanation: the OOD test used a different cabinet, and if that cabinet happened to have slightly more favorable geometry or lighting for the visual policy, the OOD improvement is an artifact of test environment selection rather than genuine robustness. This is a meaningful caveat for anyone reading the 0/Chunk robustness results as a strong endorsement — the OOD evaluation may not be a clean apples-to-apples comparison of generalization capability.