Tenma: Satellite, Robotics & UAV Routing

• Tenma is a polysemous term defining three distinct technical entities in astronomy, robotics, and wireless networking.
• In astronomy, Tenma refers to the Japanese X-ray satellite whose precise measurements of GX 339–4 establish a benchmark for inner-disk radius stability.
• In robotics and UAV routing, Tenma denotes a diffusion-transformer policy for cross-embodiment manipulation and an actor–critic method for energy-efficient multi-UAV trajectory planning.

Searching arXiv for the provided Tenma-related papers and any additional relevant context. Tenma is a polysemous term spanning distinct research contexts in astronomy, robotics, and wireless networking. In X-ray astronomy, it denotes an earlier Japanese X-ray satellite whose 1983 observations of GX 339–4 became a benchmark for later accretion-disk analyses (Shidatsu et al., 2011). In robot learning, it denotes a lightweight diffusion–transformer policy for cross-embodiment bi-manual manipulation (Davies et al., 15 Sep 2025). In wireless communications and trajectory optimization, it denotes the “Trajectory lEarNing Method based on Actor-critic,” the reinforcement-learning component of the AUTO framework for multi-UAV routing in WPT-assisted IoT systems (Dong et al., 23 Feb 2025). The term therefore does not identify a single unified concept; rather, it refers to three unrelated technical entities whose significance is defined by their respective methodological roles.

1. Astronomical usage: Tenma as an X-ray satellite

Within the astronomical literature, Tenma refers to a Japanese X-ray satellite that observed GX 339–4 during its 1983 outburst and is cited as an historical reference point for high/soft-state disk spectroscopy (Shidatsu et al., 2011). The relevant analysis states that GX 339–4 was observed three times with Tenma in 1983, and that these spectra were sufficiently detailed to be modeled with the multi-color disk (MCD) framework. This situates Tenma within early disk-spectral studies of Galactic black hole binaries in outburst.

The role of Tenma in this context is not to provide a full mission profile, but to anchor a longitudinal comparison of inferred inner-disk radii. After applying correction factors for the boundary condition and color hardening factor, the reinterpreted Tenma measurements yield inner radii of $59 \pm 1$, $59 \pm 1$, and $58 \pm 1$ km for $i=46^\circ$ and $d=8$ kpc (Shidatsu et al., 2011). These values are explicitly used as a benchmark against later observations by MAXI/GSC and Swift/XRT.

This use of Tenma is methodologically important because it supports the claim that the innermost disk radius in GX 339–4 remains stable across epochs and instruments. The cited MAXI/GSC study reports that, despite flux variability over a factor of 3 during the 2010 high/soft state, the innermost disk radius remained constant at $R_{\rm in}=61 \pm 2$ km for $i=46^\circ$ and $d=8$ kpc, and that this value is consistent with the past Tenma measurements (Shidatsu et al., 2011). The consistency across decades is taken as evidence for the standard high/soft-state interpretation in which the disk extends to the ISCO.

A central implication follows from identifying $R_{\rm in}$ with $R_{\rm ISCO}$ under a non-spinning assumption. For $59 \pm 1$0, the relation $59 \pm 1$1 together with $59 \pm 1$2 km gives a black-hole mass estimate of $59 \pm 1$3 for $59 \pm 1$4 and $59 \pm 1$5 kpc (Shidatsu et al., 2011). Combined with the optical mass function and the allowed ranges $59 \pm 1$6 kpc and $59 \pm 1$7, the allowed mass interval becomes $59 \pm 1$8 (Shidatsu et al., 2011). Tenma’s contribution is indirect but nontrivial: its agreement with later instruments strengthens confidence that the inferred radius is physical rather than a modeling or calibration artifact.

2. Tenma in studies of GX 339–4: benchmark, cross-check, and inference support

The scientific significance of Tenma in the GX 339–4 literature lies in how its measurements are recontextualized rather than in any new data reduction performed by the later paper. The reprocessed Tenma radii are expressed under the same assumptions as the MAXI/GSC analysis—specifically $59 \pm 1$9 and $58 \pm 1$0 kpc—and corrected for the same boundary and color-hardening effects (Shidatsu et al., 2011). This makes the comparison instrument-agnostic at the level of the inferred physical parameter.

The resulting cross-mission comparison can be summarized concisely.

Measurement set	Reported $58 \pm 1$1 under common assumptions	Role
Tenma, 1983 outburst	$58 \pm 1$2, $58 \pm 1$3, $58 \pm 1$4 km	Historical benchmark
MAXI/GSC, 2010 high/soft state	$58 \pm 1$5 km	Long-term monitoring reference
Swift/XRT, 2010	$58 \pm 1$6 km	Simultaneous soft X-ray check

These numerical agreements are used to reinforce three linked propositions (Shidatsu et al., 2011). First, the MCD-based radius is stable in the high/soft state despite substantial luminosity variation. Second, the disk plausibly reaches the ISCO in that state. Third, continuum-based mass and spin inferences gain credibility when the same radius scale recurs across different observatories.

The paper also notes a discrepancy with RXTE high/soft-state radii, reported as $58 \pm 1$7 and $58 \pm 1$8 km for the 1998 outburst under the same $58 \pm 1$9 and $i=46^\circ$0 assumptions, and states that the reason is unclear (Shidatsu et al., 2011). This is relevant to a common misconception: cross-instrument agreement is not perfect, and the Tenma-MAXI consistency should not be read as implying that all missions necessarily return identical radii. A more precise reading is that Tenma functions as one historically important external check within a broader, partly heterogeneous measurement record.

3. Robotic usage: Tenma as a diffusion–transformer policy

In robotics, Tenma denotes a model introduced as “a lightweight diffusion-transformer for bi-manual arm control” designed for cross-embodiment robot manipulation (Davies et al., 15 Sep 2025). It is formulated as a policy

$i=46^\circ$1

where the conditioning variables are language instructions, multiview RGB observations over a history window, and proprioceptive states over the same window, and the output is a future action sequence (Davies et al., 15 Sep 2025). The system targets real-world tabletop tasks such as food preparation, pouring, and stirring.

Architecturally, the model is organized as an encoder–decoder system with three principal components: a cross-embodiment normalizer, a Joint State–Time encoder, and a diffusion action decoder (Davies et al., 15 Sep 2025). The cross-embodiment normalizer maps heterogeneous robot state and action spaces into a fixed canonical slot structure with an activity mask and per-embodiment min–max scaling to $i=46^\circ$2. This design is intended to avoid linear memory growth with the number of embodiments and to expose a structurally consistent interface to the transformer.

The tokenization and multimodal fusion stack are also central. Tenma uses DINOv2-S ViT for RGB inputs, T5-small for language, and a 2-layer MLP with RMSNorm and GELU for proprioception, with all modalities projected to a shared token dimension of 384 (Davies et al., 15 Sep 2025). Instead of collapsing visual observations into a single token, it retains the full patch sequence. The Joint State–Time encoder then applies factorized attention across temporal, within-state, and language-conditioning axes, reducing full-attention complexity from $i=46^\circ$3 to

$i=46^\circ$4

thereby removing the dominant $i=46^\circ$5 term (Davies et al., 15 Sep 2025).

The action decoder instantiates the policy as a conditional DDPM with a DiT-style transformer. Training minimizes a diffusion noise-prediction objective,

$i=46^\circ$6

and reverse denoising uses a cosine schedule with 1000 training steps and DDIM-style inference accelerated to 7 steps (Davies et al., 15 Sep 2025). AdaLN-zero is used for diffusion-timestep conditioning, and the final decoder outputs standardized joint or Cartesian commands via a two-layer MLP. The model therefore combines sequence modeling, multimodal fusion, and generative trajectory synthesis in a single policy.

4. Cross-embodiment learning and empirical performance in robotics

The robot-learning Tenma is explicitly framed as a cross-embodiment system. Its pretraining uses a curated subset of Open X-Embodiment comprising approximately 2000 demonstrations, 20 tabletop manipulation tasks, and 7 different robot embodiments, followed by domain fine-tuning on approximately 450 human demonstrations collected on a dual-arm FAIRNO FR5 platform (Davies et al., 15 Sep 2025). The fine-tuning domain includes Pour, PickCup, PlacePan, and Stirring, with 4-view RGB at 10 Hz, synchronized end-effector state/action vectors, and roughly 10 paraphrased language prompts per trajectory.

Learning is pure imitation rather than reinforcement learning. The optimization stack uses AdamW with learning rate $i=46^\circ$7, $i=46^\circ$8, $i=46^\circ$9, $d=8$0, weight decay $d=8$1, global batch size 256, warmup steps 500, BF16 precision, dropout 0, no gradient clipping, and EMA with $d=8$2; training is performed on 8× A100 GPUs (Davies et al., 15 Sep 2025). RMSNorm, QKNorm, AdaLN-zero, and pretrained DINOv2-S and T5-small encoders are identified as stabilization choices.

The reported success rates are central to the system’s evaluation (Davies et al., 15 Sep 2025).

Model	ID average SR	Object-shift SR	Scene-shift SR
Tenma	88.95%	72.56%	81.13%
DiT-Policy	18.12%	6.09%	4.79%
Diffusion Policy	16.25%	6.64%	4.13%
Octo	11.91%	4.23%	5.59%

Task-specific in-distribution values further distinguish the architecture: Pour 93.75%, PickCup 100%, PlacePan 78.38%, and Stirring 83.67% (Davies et al., 15 Sep 2025). The paper attributes the performance gap to a combination of cross-embodiment normalization, ViT-based multi-token visual conditioning, Joint State–Time factorized attention, and a DiT decoder with AdaLN and multi-token conditioning, rather than to brute-force scaling.

Efficiency claims are comparatively restrained. Tenma is reported as a moderately sized model at approximately 110M parameters, compared with approximately 90M for DiT-Policy, approximately 50M for Diffusion Policy, and approximately 200M for Octo (Davies et al., 15 Sep 2025). On an RTX 3090 with 4 cameras, Tenma reaches 20.2 Hz inference, compared with 31.0 Hz for DiT-Policy, 13.9 Hz for Diffusion Policy, and 10.6 Hz for Octo (Davies et al., 15 Sep 2025). This suggests a speed–accuracy tradeoff rather than simple dominance along all axes.

5. Networking and optimization usage: TENMA as an actor–critic method

In wireless communications and UAV routing, TENMA stands for “Trajectory lEarNing Method based on Actor-critic” and is the reinforcement-learning algorithm that trains the Attention Trajectory Optimization Model (ATOM) in the AUTO framework (Dong et al., 23 Feb 2025). Here the term is not the name of a platform or model family but of a specific training method tailored to large-scale multi-UAV trajectory planning in WPT-assisted IoT systems.

The overall framework decomposes into ATOM as actor and TENMA as learning engine (Dong et al., 23 Feb 2025). ATOM itself is a graph-transformer-based model operating on an IoT graph

$d=8$3

where each IoTD $d=8$4 has features

$d=8$5

A graph encoder produces node-level self-attention features and a graph-level representation $d=8$6, while the trajectory decoder sequentially generates a global route sequence $d=8$7 that can later be segmented into multiple UAV routes (Dong et al., 23 Feb 2025).

The decoder state is defined as

$d=8$8

with alignment weights

$d=8$9

context

$R_{\rm in}=61 \pm 2$0

and masked next-node probabilities

$R_{\rm in}=61 \pm 2$1

followed by

$R_{\rm in}=61 \pm 2$2

These expressions define the actor’s sequential route-construction mechanism (Dong et al., 23 Feb 2025).

TENMA then trains this actor with Monte Carlo rollouts. For sampled trajectories $R_{\rm in}=61 \pm 2$3, the reward is defined as

$R_{\rm in}=61 \pm 2$4

which corresponds to negative total flight distance and is aligned with minimizing flight energy (Dong et al., 23 Feb 2025). The policy gradient uses a critic baseline: $R_{\rm in}=61 \pm 2$5 while the critic is trained by

$R_{\rm in}=61 \pm 2$6

The defining methodological claim is that the critic tracks the real system reward, thereby reducing policy-gradient variance and stabilizing large combinatorial optimization (Dong et al., 23 Feb 2025).

6. Optimization problem, experiments, and comparative results for TENMA in UAV routing

The TENMA-based AUTO framework addresses a decomposed optimization problem in WPT-assisted IoT systems (Dong et al., 23 Feb 2025). The full problem $R_{\rm in}=61 \pm 2$7 minimizes total UAV energy under association, return, service, WPT/data-upload, battery, and storage constraints. It is split into a time-allocation problem $R_{\rm in}=61 \pm 2$8, solved via KKT with closed-form $R_{\rm in}=61 \pm 2$9 and $i=46^\circ$0, and a trajectory-planning problem $i=46^\circ$1, which is a Capacitated Vehicle Routing Problem (CVRP) minimizing $i=46^\circ$2 (Dong et al., 23 Feb 2025). ATOM parameterizes $i=46^\circ$3, while TENMA learns parameters that approximately solve $i=46^\circ$4 through reward maximization.

The reported training setup uses a simulated $i=46^\circ$5 area with $i=46^\circ$6 IoTDs and $i=46^\circ$7 uniformly sampled in $i=46^\circ$8 MB (Dong et al., 23 Feb 2025). Hyperparameters include $i=46^\circ$9 self-attention layers, $d=8$0 heads, feed-forward hidden dimension 512, actor learning rate $d=8$1, critic learning rate $d=8$2, decay 0.98, 500 training epochs, experience pool capacity 10,000, and batch size 100 (Dong et al., 23 Feb 2025). The paper reports stable convergence.

Comparative energy results are especially important because they isolate the effect of TENMA relative to alternative RL trainers (Dong et al., 23 Feb 2025).

Trajectory designer	Reinforce	AC	A2C	TENMA
PN	287.88	287.42	285.04	285.83
GNN	293.46	292.73	290.31	290.72
GPN	275.48	273.25	272.85	274.46
ATOM	266.51	265.17	264.84	264.32

For ATOM specifically, TENMA yields the lowest average energy cost at 264.32 Wh (Dong et al., 23 Feb 2025). Additional results show that AUTO, i.e. ATOM plus TENMA, consistently attains the lowest energy cost under battery-capacity variation from 1700 to 2550 mAh, outperforming PN, GPN, and GNN (Dong et al., 23 Feb 2025). In a field case at Yuelu Mountain, Hunan, China, involving 125 air quality sensors and a DJI Matrice 100 with WPT hardware, the planning time is reported as 1.04 s, AUTO produces 7 UAV trajectories, and ATOM with multi-head self-attention achieves 72.73 Wh, the lowest energy among compared schemes (Dong et al., 23 Feb 2025). These results position TENMA as a specialized RL trainer for structured route generation rather than a general-purpose actor–critic algorithm.

7. Comparative interpretation and disambiguation

The three uses of Tenma are unrelated in origin and technical substance. The astronomical Tenma is a historical observational platform used as a cross-era calibration reference in black-hole accretion studies (Shidatsu et al., 2011). The robotics Tenma is a policy architecture for multimodal cross-embodiment imitation learning with diffusion transformers (Davies et al., 15 Sep 2025). The networking TENMA is a learning algorithm inside a graph-transformer-plus-RL pipeline for CVRP-like multi-UAV routing under WPT-assisted IoT constraints (Dong et al., 23 Feb 2025).

A concise disambiguation is therefore useful.

Usage	Domain	Technical role
Tenma	X-ray astronomy	Japanese X-ray satellite used as historical benchmark for GX 339–4 disk-radius measurements
Tenma	Robot learning	Lightweight diffusion–transformer policy for cross-embodiment bi-manual manipulation
TENMA	Wireless networking / RL	“Trajectory lEarNing Method based on Actor-critic” for training ATOM in AUTO

A common misconception would be to treat these references as variants of a shared research program. The available evidence does not support that reading. The shared name is nominal only. The astronomy usage concerns observational inference of $d=8$3, black-hole mass, and spin baselines (Shidatsu et al., 2011). The robotics usage concerns multimodal tokenization, factorized attention, diffusion decoding, and success-rate generalization under object and scene shifts (Davies et al., 15 Sep 2025). The networking usage concerns graph-structured policy optimization, critic-baseline variance reduction, and energy-efficient multi-UAV trajectory design (Dong et al., 23 Feb 2025).

What unifies the term at most is a functional pattern: in each field, Tenma occupies a structurally central role. In astronomy, it anchors validation across missions; in robotics, it integrates heterogeneous modalities and embodiments into a single policy; in UAV routing, it stabilizes learning for large combinatorial action spaces. This suggests a purely editorial generalization: “Tenma” often appears not as a peripheral label but as the name of the component or instrument on which downstream claims materially depend.