Learned Perceptive Forward Dynamics Model for Safe and Platform-aware Robotic Navigation
1. Paper overview
In one sentence: Predicting platform-specific motion and failure risk makes MPPI navigation more effective on tested rough terrain, while retaining dependencies on training coverage and planner tuning. e02e04e07e12e13e15
| At a glance | What to know |
|---|---|
| Research problem | Source description A commanded velocity is an unreliable proxy for actual movement when a quadruped climbs, slips, collides or cannot traverse an obstacle. Handcrafted traversability costs struggle to express platform-specific capabilities. The paper learns command-conditioned future poses and failure risk so a sampling planner can reason about these interactions using a compact local representation. e02e03 |
| Core mechanism | |
| A key reported result | MPPI navigation in simulated 2D and 3D environments: Ours: 88.33% in 2D; 73.75% in 3D. Successful 3D paths average 3.93 m and 8.68 s. Goal success (%); successful-path mean length (m) and time (s). Table III; synthetic-only FDM. Baseline planners are separately tuned, and the Kim model is trained only in 2D for this planning comparison. Success: Kim 78.33%/48.75%, heuristics 82.50%/33.13% in 2D/3D. Kim's successful 3D paths average 7.20 m and 18.60 s. The 3D success advantage over Kim is 25.00 percentage points, calculated from the table. Unequal sampling settings and timeouts prevent a strictly controlled model-only attribution; successful-path means condition on different subsets. e13e18 |
| Reading caution | Source description Generalization is restricted by training geometry and locomotion capability; novel confined spaces, ice or deep mud may fail. Real data excludes damaging collisions, leaving failure-label transfer unresolved. Fast-moving objects, multiple dynamic agents and social norms are not tested. e15 |
Core contributions
Figure 2. Training connects commanded motion to observed outcomes through a residual dynamics model. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read from left to right. Simulation and real trajectories enter the replay buffer; sampling separates commands, observation history, the current height scan and future-state labels. In the blue model block, a history GRU and a height-scan encoder produce context for forward prediction. Candidate commands pass through an action encoder, while the upper command line also reaches trajectory integration. That bypass matters: Section V-B explains that the pose head predicts corrections to intended velocities, which are added before integration. The lower prediction branch estimates failure probability. Follow the bottom future-state line to the loss: these are training targets, not future observations supplied during deployment. e03e04e06e20
What it supports. The architecture makes the prediction conditional on both the proposed motion and what the robot can currently perceive. Its learned correction can capture tracking error and terrain interactions that constant-velocity integration misses. The two output branches support different planning questions: where a command sequence leads and whether it is risky.
Where the evidence stops. The graphic labels proprioception as m1…6, but Table I and Section V-B specify m1…7, including previous joint actions. This guide follows the table. The appendix module listing also does not fully resolve the context-to-recurrence wiring described in the main text.
2. Motivation
2.1 The problem and the proposed response
A commanded velocity is an unreliable proxy for actual movement when a quadruped climbs, slips, collides or cannot traverse an obstacle. Handcrafted traversability costs struggle to express platform-specific capabilities. The paper learns command-conditioned future poses and failure risk so a sampling planner can reason about these interactions using a compact local representation. e02e03
2.2 What this reading follows
A velocity command tells a robot what to attempt; terrain and the locomotion controller determine what actually happens. This paper learns that gap from trajectories. Its forward dynamics model combines a local height scan with recent robot measurements, predicts where candidate commands will lead, and estimates failure risk. A separate sampling planner uses these predictions to select the next command. The six visuals below trace the training pipeline, the planning loop, the main comparisons and two ablations. Read the simulation navigation results separately from real-world prediction improvements: the latter support adaptation of motion estimates, but do not quantify real-world navigation safety. e02e04e07e12e13e15
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The core WAM/Dual-system/Other mechanisms/Outside quadrants classification is supported: an action-conditioned forward model predicts poses and risk, while a separate sampling controller selects actions. This is neither inverse dynamics nor joint future-and-action generation. Navigation is explicit; '3D multiview modeling' is only a loose descriptor because the learned input is a height scan, not an explicit multiview reconstruction objective. e03e04e07
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Predict the motion that follows a command
Imagine offering the same velocity sequence to a robot facing a wall and to one facing traversable stairs. Constant-velocity integration gives the same nominal motion in both cases. The FDM first reads the local height scan and recent robot measurements, then conditions a forward recurrent network on the candidate sequence. Its correction head estimates how followed velocities differ from intended commands, and its risk head estimates failure along the prediction. The training target is expressed in the robot frame at the start of the sample. Ten predictions spaced by 0.5 seconds provide a five-second forecast, while historical measurements are spaced by 0.05 seconds. These are distinct temporal roles, motivating separate history and forward recurrent units. The model predicts pose and risk without reconstructing future camera images. e03e04e05e08
Figure 3. The planner searches predicted outcomes and repeatedly feeds its solution back into candidate generation. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the initial solution in the green block and follow the perturbation arrow down to the candidate action distribution. The candidate sequences enter the blue FDM along with observation o_t. Its predicted poses and risk flow into the yellow reward block, which combines goal progress and risk assessment. The upward arrow returns the preferred actions to the next iteration. The red-to-blue scale on the right marks low-to-high reward, not a measured success probability. Figure 3 emphasizes maximum-reward selection; Equations (6)–(7) additionally describe reward-weighted perturbation updates. After the search, the receding horizon system executes the next command and observes the robot again. e07e19
What it supports. Action selection belongs to MPPI, while the learned network evaluates consequences. This separation supports changing planning objectives without training a new navigation policy. It also explains the paper's main taxonomy placement: command-conditioned forward prediction is used inside an external controller, rather than jointly generating future states and actions.
Where the evidence stops. Figure 3 omits the reward weights shown in Equation (13). More seriously, maximization coexists with a distance expression lacking an explicit minus sign, and the risk weight appears in both Equations (13) and (15). The intended goal-seeking, risk-penalizing behavior is clear; exact signs and scaling remain unresolved.
5.2 Separate motion adaptation from proof of safety
Simulation provides the failures that would be too costly to collect on hardware. The training loss therefore supervises both poses and risk, and keeps target poses stationary after a failure. Safe real trajectories add information about slipping, vegetation and sensor behavior that rigid-body simulation misses. Section VI-D reports lower prediction error after this fine-tuning on new datasets in similar environments. Reader analysis: better motion prediction and calibrated failure probability are different claims. Safe real data can correct motion bias without showing whether simulated collision labels transfer faithfully. This matters because MPPI may exploit a falsely optimistic prediction during search. The paper acknowledges the absent real collision demonstrations, and its physical deployments are qualitative navigation evidence. They should not be assigned the simulation success percentages or treated as a quantified safety guarantee. e05e06e12e14e15
5.3 Locate where the planner's advantage comes from
MPPI uses the learned forecast as an evaluator. It perturbs candidate command sequences, scores predicted terminal progress and risk, refines its solution, and executes the next command before replanning. The headline planning comparison tests this complete arrangement against separately tuned alternatives. Reader analysis: Table VI is more specific about mechanism because it removes the risk reward from the proposed planner, whereas Table III changes models and planner settings together. The ablation's higher success with longer time supports a cautious-motion tradeoff in the tested setup. It does not establish that every part of the full comparison comes from the risk head. A reproduction must also distinguish the 2048-trajectory onboard configuration from the appendix's simulation baseline settings, and resolve the printed reward signs before using the equations as executable specifications. e07e09e13e17e18
5.4 Training and inference
During training
Ten future states at 0.5-second intervals give a five-second horizon; ten historical states use 0.05-second intervals. Synthetic training uses 15 rounds, each collecting 80k samples from 10k parallel environments, followed by eight update episodes with batch size 2048. AdamW uses a scheduler and weight decay. The reported run takes approximately eight hours on one RTX 4090. e08
Pose MSE includes sine/cosine heading encoding; risk uses binary cross-entropy. Failure trajectories hold target poses constant, with an additional stop loss. Observation augmentation includes noise, missing height patches and camera occlusions. Fine-tuning mixes real and synthetic data at a small constant learning rate; no frozen-module stage is specified. e04e05e06e08
During inference
MPPI perturbs candidate action sequences, evaluates FDM poses and risks, and refines the solution before executing the next command in a receding horizon loop. The goal term concerns terminal distance; a thresholded risk term also considers neighboring paths. Reported onboard operation uses a Jetson Orin AGX, 2048 trajectories, 7 Hz planning and 40.6 ms model inference per iteration. e07e09
5.5 Implementation flow
- Collect platform-specific trajectories
Simulation supplies risky interactions; an expert collects safe real trajectories. Replay buffers yield random sub-trajectories with past observations and future labels, and poses are expressed in the robot frame at the sampling time. Early synthetic commands use time correlation; later collection partly uses the current FDM planner. e05
- Encode observations and candidate actions
Section V-B describes a history GRU and height-scan CNN whose embeddings condition a forward GRU. An MLP encodes candidate commands. Separate heads predict velocity corrections and risk; corrected commands are integrated into poses. Figure 2 shows the command bypass into integration and future labels entering the training loss. e04
- Represent embodiment through training
Individual FDMs are trained for ANYmal, ANYmal-On-Wheels and Barry with different locomotion policies. Their differing predictions reflect platform and controller experience, rather than demonstrating one shared model transferring unchanged among robots. e11e19
6. Experiments & results
A perceptive forward dynamics model predicts how a particular robot and locomotion policy will respond to velocity commands on nearby terrain. A separate MPPI planner searches those predictions for useful, low-risk motion. Simulation comparisons support improved navigation, while real-world data improves trajectory prediction; neither establishes universal safety.
6.1 Read the original evidence
Table II. Prediction quality improves in complex terrain, but collision recall and rare-event behavior need separate attention. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read one environment block at a time. Position offset measures final predicted-location error in metres; lower is better. The remaining columns evaluate failure classification, with a risky action sequence treated as positive. Precision asks whether a predicted failure is correct; recall asks how many actual failures are detected. The displayed precision, recall and accuracy values use a percentage scale, while F1 uses a zero-to-one scale. Dashes for constant velocity indicate absent failure predictions. Section VI-B uses 50k evaluation samples per environment and trains the learned baseline on the same data, while preserving its original architecture, objective and training hyperparameters. e10e15
What it supports. In 3D terrain, the proposed model reports 0.28 ± 0.35 m position offset, compared with 0.44 ± 0.45 m for Kim and 0.99 ± 1.06 m for constant velocity. F1 rises from Kim's 0.80 to 0.85, while recall is slightly lower: 86.51% versus 86.62%.
Where the evidence stops. Planar failures occur in only 0.042% of samples: 98.32% accuracy therefore coexists with F1 0.10. High accuracy is not a safety guarantee. The table does not define ± as a confidence interval, and its 2D accuracy of 89.13% contradicts the conclusion's stated minimum of 89.20%.
Table III. Improved predictions translate into higher goal-reaching success under the paper's separately tuned planning configurations. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with Success, then distinguish the Suc. and All subcolumns under mean path length (MPL) and mean path time (MPT). Suc. includes trajectories that reach the goal; All also includes failures. Early collisions can shorten averages, whereas timeouts and circling can lengthen them. Comparisons therefore need both the success rate and the appropriate subset. The 3D block is particularly relevant because stairs and ramps challenge the baseline's horizontal sensing and the heuristic cost map. Unlike Table II, the Kim FDM here is trained only on 2D terrain. Appendix F also reports changes to its candidate population, risk weight and timeout. e13e18e14
What it supports. In 3D simulation, success is 73.75% for the proposed planner, 48.75% for Kim and 33.13% for heuristics. The advantage over Kim is 25.00 percentage points, calculated from these entries. Among successful trajectories, the proposed method reports 3.93 m and 8.68 s versus Kim's 7.20 m and 18.60 s.
Where the evidence stops. These are system comparisons with unequal planner configurations, not an isolated estimate of the FDM's causal effect. Trial counts, seed variability and confidence intervals are not supplied. Successful-path means compare different subsets, and the table does not quantify physical-robot navigation success.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Five-second dynamics prediction in simulated 3D terrain Section VI-B: 50k evaluation samples per environment; Kim et al. is trained on the same data while retaining its original architecture, loss and hyperparameters. | Ours: 0.28 ± 0.35 m and F1 0.85. Final position offset (m); collision F1 | Kim et al.: 0.44 ± 0.45 m and F1 0.80; constant velocity: 0.99 ± 1.06 m, without collision predictions. Lower position error accompanies better F1, but recall is 86.51% versus Kim's 86.62%. Table II does not identify its ± statistic as a confidence interval. e10 |
| MPPI navigation in simulated 2D and 3D environments Table III; synthetic-only FDM. Baseline planners are separately tuned, and the Kim model is trained only in 2D for this planning comparison. | Ours: 88.33% in 2D; 73.75% in 3D. Successful 3D paths average 3.93 m and 8.68 s. Goal success (%); successful-path mean length (m) and time (s) | Success: Kim 78.33%/48.75%, heuristics 82.50%/33.13% in 2D/3D. Kim's successful 3D paths average 7.20 m and 18.60 s. The 3D success advantage over Kim is 25.00 percentage points, calculated from the table. Unequal sampling settings and timeouts prevent a strictly controlled model-only attribution; successful-path means condition on different subsets. e13e18 |
| Real-world trajectory prediction after fine-tuning New datasets in environments similar to the forest, snow and pavement training deployments; Figure 8 presents prediction steps 4 and 9. | Forest 34.38%; snow 30.55%; pavement 30.30%. Reported relative reduction in mean position error | Fine-tuned FDM versus the synthetic-only FDM; constant velocity is also plotted. Kim is unavailable because these datasets lack 2D LiDAR. These are trajectory-error reductions, not real-world navigation success gains. Exact split sizes and uncertainty on the reductions are not provided. e12 |
| Input-modality ablation in simulated 3D terrain Appendix D, Table V; compare variants within this ablation table. | Full model: 0.27 ± 0.32 m; F1 0.88. Position offset (m); collision F1 | Without height scan: 0.41 ± 0.41 m; F1 0.80. Without past-state observations: 0.27 ± 0.32 m; F1 0.87. Height information affects both outcomes; state history primarily affects failure estimation here. Full-model values differ from Table II and should not be pooled with it. e16 |
| Risk-reward ablation in simulated 3D planning Appendix D, Table VI; the prose identifies the intervention as removing the planner's failure-risk reward. | With risk: 73.75%, 8.68 s. Success (%); mean path time (s) | Without risk: 69.17%, 5.74 s. The calculated success difference is 4.58 percentage points, with slower motion. The table's broader row label 'w/o Failure Estimation' does not establish that the FDM risk head was retrained or removed. e17 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table V. Removing height scans damages both pose prediction and risk assessment; other modalities have environment-dependent effects. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each block starts with the complete model, followed by variants without state observations, proprioceptive observations or height scans. Make comparisons within a block and against this table's own full-model row. For the 3D block, removing state observations preserves the displayed position offset but slightly lowers F1. Removing the height scan raises position error and lowers F1 more substantially. The planar block shows a different sensitivity: losing proprioception strongly damages position prediction. Read the failure columns together, since higher recall can accompany poorer precision. These ablations concern learned prediction performance; they do not by themselves measure how each missing modality changes closed-loop navigation. e16e10
What it supports. For the 3D ablation, the full model has 0.27 ± 0.32 m offset and F1 0.88. Without height scans, these become 0.41 ± 0.41 m and 0.80. Without state observations, offset remains 0.27 ± 0.32 m and F1 becomes 0.87, showing a smaller effect in this environment.
Where the evidence stops. The full-model scores differ from Table II, so do not pool the two tables. Appendix D says history improves F1 by up to 0.06, but the planar rows show 0.37 versus 0.28, a 0.09 difference. The table gives no uncertainty across retraining runs.
Table VI. Adding the risk reward trades faster completion for a higher observed goal-reaching rate. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the two rows inside each environment before looking across environments. The first row uses the proposed planner; the second is labeled 'w/o Failure Estimation'. Appendix D narrows that label to removing the planner's failure-risk reward. It does not establish that the prediction network loses its risk head or is retrained. Read Success as a percentage, MPL as metres and MPT as seconds. In both environment blocks, the risk-aware row has higher success and longer reported time, while path lengths remain close. This is the paper's most direct diagnostic of whether explicitly scoring predicted risk helps the planner beyond its pose predictions. e17e13
What it supports. In 3D, adding risk changes success from 69.17% to 73.75%, a calculated gain of 4.58 percentage points. Reported mean time changes from 5.74 s to 8.68 s, with lengths of 3.96 m and 3.93 m. In 2D, success similarly rises from 85.86% to 88.33%.
Where the evidence stops. Unlike Table III, this table does not label successful versus all-path aggregation; its full-model length/time entries match Table III's successful-path values. No confidence intervals are given. The rounded 2.5% and 4.6% prose changes should be read as percentage-point differences, not relative improvements.
7. Analysis & limitations
7.1 What the evidence leaves open
Generalization is restricted by training geometry and locomotion capability; novel confined spaces, ice or deep mud may fail. Real data excludes damaging collisions, leaving failure-label transfer unresolved. Fast-moving objects, multiple dynamic agents and social norms are not tested. e15
High accuracy is insufficient evidence of safety: Table II's planar failure rate is only 0.042%, while the proposed model's planar F1 is 0.10 despite 98.32% accuracy. The conclusion's claimed minimum accuracy of 89.20% also overlooks the table's 2D value of 89.13%. e10e15
7.2 Questions for discussion
- How much of Table III's advantage persists when planner population, action distribution and timeouts are held equal?
- Can safe real-world data improve failure-risk calibration when damaging failures are absent from that data?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires the platform simulator, matching locomotion policy, observation construction and augmentation, replay sampling, and labeled real trajectories. Appendix C lists a 1.16-million-parameter model; Appendix E specifies terrain composition and distinguishes FDM collection from domain randomization used in locomotion-policy training. e05e20
Resolve source ambiguities before implementation: Figure 2 labels six proprioceptive groups while Table I lists seven; Section V-B describes hidden-state initialization, but Appendix C's module listing does not specify the forward wiring. Equation (11) conditions on un-tilded risk despite prose about predicted failure. Reward maximization, the distance expression and repeated risk weight in Equations (13)–(15) leave signs/scaling unclear. e03e04e06e07e20
The text leaves exact learning rates, loss weights, fine-tuning mixture, software versions, dataset split sizes and evaluation seed counts incomplete or unstated. Proposed checks should match command sampling and separately audit risk calibration and navigation success. Platform changes require updated simulation, locomotion and input layers. e08e12e13e18e19
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Repeat the height-scan ablation on paired terrain and command samples
Reader-proposed check, not performed: train full and no-height-scan variants on the same trajectory splits, command distribution, augmentation policy and training budget, with multiple matched seeds. Evaluate both on identical five-second command sequences in held-out walls, ramps and stairs. Report final position error, failure precision/recall/F1, and errors separately for collision and non-collision cases, following the motivation of Figure 13. Control for changed parameter count when removing the encoder, and document that choice. The mechanism predicts a persistent advantage for the full model on geometry-dependent outcomes. If that gap disappears under matched data and retraining, the claimed necessity of height information would need a narrower interpretation. e04e08e10e16e21
Check 2: Isolate the planner's risk reward with a fixed learned model
Reader-proposed check, not performed: keep one trained FDM checkpoint, candidate count, initial poses/goals, action limits, perturbation seeds, neighbor rule and timeout fixed. Toggle only the planner's risk contribution. Before navigation, verify that the implemented reward prefers a closer safe endpoint and penalizes an otherwise identical risky trajectory; document how the printed sign/scaling ambiguity was resolved. Run paired simulation trials and report success, collision frequency, timeouts, and length/time for both successful and all paths. Reproducing higher success with risk at the cost of longer time would support Table VI's interpretation. An effect that vanishes when sampling and timeout are matched would weaken attribution to explicit risk scoring. e07e13e17e18e19
8.3 Reading coverage
Visual audit: Original PDF pages 1–10 and 13–18 were rendered and visually inspected, covering the title/authors/version, every figure (1–14), every table (I–VI), method equations, training and onboard hardware, evaluation protocols, limitations, architecture listing and platform adaptation. Pages 11–12 contain acknowledgments/references and were read in the complete text pass. All seven supplied text chunks were read individually. Each of the six final crops was viewed; the risk-ablation crop was tightened to remove caption spillover and then viewed again. Figure 2's proprioceptive label was checked against Table I and Section V-B; the MPPI arrows and maximum-reward label were checked against Equations (4)–(7) and (13)–(15). Source discrepancies are disclosed in the corresponding cautions. Separate supplements, linked code, datasets and videos remain outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 13, 14, 15, 16, 17, 18. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- PDF pp. 1–3: abstract, introduction and related work
- PDF pp. 3–6: preliminaries, problem statement and methodology, Sections III–V
- PDF pp. 6–10: experiments VI-A–F, limitations and conclusions
- PDF pp. 11–12: acknowledgments and references
- PDF pp. 13–18: Appendix A–H, including nomenclature, action sampling, architecture listing, design ablations, terrain design, expanded experiments and platform adaptation
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The extraction limitation was addressed by inspecting original PDF pages 1–10 and 13–18, including every figure and table. Reference pages 11–12 were read as text.
- The reviewed artifact is arXiv:2504.19322v2, dated 29 April 2025 and marked accepted for RSS 2025. Its title and all four authors match the catalog. The catalog submission date is 27 April 2025. The acquisition snapshot records separate RSS identity verification; this reading did not inspect the proceedings artifact and does not establish byte or content equivalence between editions.
- No separate supplements, linked code, datasets or project videos were inspected, and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block, arXiv margin and acceptance note
The exact catalog title and authors Pascal Roth, Jonas Frey, Cesar Cadena and Marco Hutter appear. Affiliations are ETH Zurich, NVIDIA and Max Planck Institute for Intelligent Systems. The margin identifies arXiv:2504.19322v2, 29 April 2025; the footnote marks acceptance for RSS 2025.
Go to primary source ↓e02PDF pp. 1–3, Abstract, Section I and Section II-B
The paper motivates platform-dependent dynamics and traversability, combining learned forward prediction with sampling-based planning instead of only handcrafted traversability assessment.
Go to primary source ↓e03PDF p. 3, Sections III-A and IV-A, Equation (3); p. 4, Table I
The n-step model conditions on observations and candidate actions; states comprise planar pose and risk, actions comprise x/y/yaw velocity. Table I enumerates seven proprioceptive groups and a height map.
Go to primary source ↓e04PDF p. 4, Figure 2 and caption; p. 5, Section V-B
History GRU and height-scan CNN condition forward recurrence; an action encoder and two heads provide velocity corrections and failure probabilities. Commands bypass to integration and future states supervise loss. Figure 2 labels m1…6, whereas the text uses m1…7. Height scans include missing-patch and visibility augmentation.
Go to primary source ↓e05PDF pp. 4–5, Section V-A; p. 13, Appendix B; p. 15, Appendix E
Trajectory replay sampling uses base-frame poses, synthetic risky maneuvers and safe real data. Command collection shifts from correlated random sampling to partial planner sampling. Real pose labels fuse RTK-GNSS, IMU and total-station estimates. Terrain design is randomized; Appendix E says FDM collection itself did not apply domain randomization, unlike locomotion-policy training.
Go to primary source ↓e06PDF p. 5, Section V-C, Equations (9)–(12)
Training combines pose MSE, heading sine/cosine encoding, risk BCE and a weighted stop loss. Targets stop at failure. Equation (11) writes its risk condition without a tilde, while the preceding prose discusses predicted failure.
Go to primary source ↓e07PDF p. 3, Section III-B, Equations (4)–(7); p. 4, Figure 3; p. 6, Section V-D, Equations (13)–(15); p. 1, Figure 1 caption
MPPI uses candidate rewards, Gaussian perturbations and weighted refinement; figures emphasize maximum-reward selection. Execution is receding horizon. The goal term uses terminal distance, and risk considers neighboring paths. Maximization is explicit, but negative penalty signs are not explicit and the risk coefficient occurs both outside and inside the risk term.
Go to primary source ↓e08PDF p. 7, Section VI, Model Training paragraph
Ten predictions at 0.5 s and ten historical samples at 0.05 s; 15 rounds of 80k samples from 10k environments, eight update episodes, batch 2048, AdamW and approximately eight hours on one RTX 4090. Real fine-tuning uses a small constant learning rate. VI-A–C/E/F use synthetic-only models; VI-D uses fine-tuning.
Go to primary source ↓e09PDF p. 6, Section VI, Experimental Setup paragraph
IsaacLab supports simulation; real deployment uses ANYmal. Onboard Jetson Orin AGX planning is reported at 7 Hz with 2048 trajectories and 40.6 ms FDM inference per iteration.
Go to primary source ↓e10PDF p. 7, Section VI-B; p. 8, Table II and caption, especially 3D, 2D and Plane rows
Prediction evaluation uses 50k samples per environment and the same training data for Kim's baseline. In 3D, ours/Kim/constant-velocity offsets are 0.28±0.35/0.44±0.45/0.99±1.06 m; ours/Kim F1 is 0.85/0.80 and recall 86.51/86.62%. Plane failures are reported at 0.042%; ours has 98.32% accuracy and 0.10 F1. Ours' 2D accuracy is 89.13%.
Go to primary source ↓e11PDF p. 8, Section VI-C and Figure 7
Separate FDMs are trained for ANYmal, ANYmal-On-Wheels and Barry; Barry's robust and quiet locomotion policies produce distinct capabilities. The wheeled platform uses an extended height scan.
Go to primary source ↓e12PDF pp. 8–9, Section VI-D and Figure 8
Real fine-tuning uses pavement, snow and forest data; evaluation uses new datasets in similar environments. Reported mean-error reductions are forest 34.38%, snow 30.55% and pavement 30.30%. The plot compares steps 4 and 9. Missing 2D LiDAR prevents evaluating Kim on these datasets.
Go to primary source ↓e13PDF p. 9, Section VI-E, Table III and caption
Success rates for ours/Kim/heuristics are 88.33/78.33/82.50% in 2D and 73.75/48.75/33.13% in 3D. Successful 3D paths for ours are 3.93 m and 8.68 s; Kim gives 7.20 m and 18.60 s. Suc. and All columns use different trajectory subsets. Kim is trained only on 2D for planning.
Go to primary source ↓e14PDF p. 9, Section VI-F; p. 1, Figure 1; p. 10, Figure 9; p. 17, Figure 14
Qualitative planning and prediction illustrations show real and simulated terrain interactions. Long-range goals are projected into perception range during deployment; these illustrations do not provide a real-world navigation success-rate benchmark.
Go to primary source ↓e15PDF p. 10, Sections VII–VIII
Limitations include training-domain geometry, locomotion capabilities, missing real collision data, planner tuning and untested social/dynamic-agent settings. The conclusion states a minimum failure-estimation accuracy of 89.20%, which differs from Table II's 2D row.
Go to primary source ↓e16PDF p. 14, Appendix D; p. 15, Table V, all rows
Input ablations compare state history, proprioception and height scans. Full versus no-height 3D offsets are 0.27±0.32 versus 0.41±0.41 m and F1 0.88 versus 0.80. No-state 3D offset stays 0.27±0.32 m with F1 0.87. In Plane, full/no-state F1 is 0.37/0.28, exceeding the prose's claimed maximum 0.06 gain.
Go to primary source ↓e17PDF p. 14, Appendix D, final paragraph; p. 15, Table VI and caption
The prose specifies removal of the planner's failure-risk reward. With/without success is 88.33/85.86% in 2D and 73.75/69.17% in 3D. Corresponding times are 9.23/7.73 s and 8.68/5.74 s. The row instead says 'w/o Failure Estimation'; no risk-head retraining is specified.
Go to primary source ↓e18PDF p. 16, Appendix F, Expanded Planning Experiments Results
Kim's planning population is halved to 256, risk reward weight is reduced tenfold and timeout is increased. The heuristic planner increases the neighboring filter from three to four. These are baseline-specific settings, distinct from the onboard deployment configuration.
Go to primary source ↓e19PDF p. 18, Appendix H
New platforms require changing simulation, locomotion policy and input layers, with possible command-sampling adjustments. Architecture/hyperparameter reuse is an assumption. Noise, correlation, reward scaling and allowed action space affect MPPI behavior.
Go to primary source ↓e20PDF pp. 13–14, Appendix C, parameter-count statement and module listing
The appendix reports 1.16 million parameters and lists CNN, MLP and two GRU modules. The recurrence is listed as GRU(596,128), and both prediction heads begin with 1280 input features; the printed module listing does not expose full forward wiring despite the hidden-initialization description in Section V-B.
Go to primary source ↓e21PDF p. 16, Appendix G and Figure 12; p. 17, Figure 13 and caption
Expanded dynamics diagnostics separate collision and non-collision errors. Figure 13 displays up to the 95th percentile and attributes high non-collision errors of the baseline in complex terrain to false collision predictions.
Go to primary source ↓8.5 Primary sources
Learned Perceptive Forward Dynamics Model for Safe and Platform-aware Robotic Navigation ↗
PDF · 12,109 extracted words
Source fingerprint
41c3267f317e235076d3035e3cc3829c92e77f4b70e0fb312085ebd0d4e261f4